Effective C: An Introduction to Professional C Programming 1718501048, 9781718501041

A detailed introduction to the C programming language for experienced programmers. The world runs on code written in th

4,238 930 20MB

English Pages 240 [274] Year 2020

Report DMCA / Copyright

DOWNLOAD FILE

Polecaj historie

Effective C: An Introduction to Professional C Programming
 1718501048, 9781718501041

Table of contents :
Brief Contents
Contents in Detail
Foreword by Pascal Cuoq
Foreword by Ollie Whitehouse
Acknowledgments
Introduction
A Brief History of C
The C Standard
The CERT C Coding Standard
Who This Book Is For
What’s in This Book
Chapter 1: Getting Started with C
Developing Your First C Program
Compiling and Running Your Program
Preprocessor Directives
The main Function
Checking Function Return Values
Formatted Output
Editors and Integrated Development Environments
Compilers
GNU Compiler Collection
Clang
Microsoft Visual Studio
Portability
Implementation-Defined Behavior
Unspecified Behavior
Undefined Behavior
Locale-Specific Behavior and Common Extensions
Summary
Chapter 2: Objects, Functions, and Types
Objects, Functions, Types, and Pointers
Declaring Variables
Swapping Values (First Attempt)
Swapping Values (Second Attempt)
Scope
Storage Duration
Alignment
Object Types
Boolean Types
Character Types
Numerical Types
void Types
Function Types
Derived Types
Pointer Types
Arrays
Structures
Unions
Tags
Type Qualifiers
const
volatile
restrict
Exercises
Summary
Chapter 3: Arithmetic Types
Integers
Padding and Precision
The Header File
Declaring Integers
Unsigned Integers
Signed Integers
Integer Constants
Floating-Point
Floating-Point Types
Floating-Point Arithmetic
Floating-Point Values
Floating-Point Constants
Arithmetic Conversion
Integer Conversion Rank
Integer Promotions
Usual Arithmetic Conversions
An Example of Implicit Conversion
Safe Conversions
Summary
Chapter 4: Expressions and Operators
Simple Assignment
Evaluations
Function Invocation
Increment and Decrement Operators
Operator Precedence and Associativity
Order of Evaluation
Unsequenced and Indeterminately Sequenced Evaluations
Sequence Points
sizeof Operator
Arithmetic Operators
Unary + and – Operators
Logical Negation Operator
Multiplicative Operators
Additive Operators
Bitwise Operators
Complement Operator
Shift Operators
Bitwise AND Operator
Bitwise Exclusive OR Operator
Bitwise Inclusive OR Operator
Logical Operators
Cast Operators
Conditional Operator
_Alignof Operator
Relational Operators
Compound Assignment Operators
Comma Operator
Pointer Arithmetic
Summary
Chapter 5: Control Flow
Expression Statements
Compound Statements
Selection Statements
The if Statement
The switch Statement
Iteration Statements
The while Statement
The do...while Statement
The for Statement
Jump Statements
The goto Statement
The continue Statement
The break Statement
The return Statement
Exercises
Summary
Chapter 6: Dynamically Allocated Memory
Storage Duration
The Heap and Memory Managers
When to Use Dynamically Allocated Memory
Memory Management Functions
The malloc Function
The aligned_alloc Function
The calloc Function
The realloc Function
The reallocarray Function
The free Function
Memory States
Flexible Array Members
Other Dynamically Allocated Storage
The alloca Function
Variable-Length Arrays
Debugging Allocated Storage Problems
Dmalloc
Safety-Critical Systems
Exercises
Summary
Chapter 7: Characters and Strings
Characters
ASCII
Unicode
Source and Execution Character Sets
Data Types
Character Constants
Escape Sequences
Linux
Windows
Character Conversion
Strings
String Literals
String-Handling Functions
and
Annex K Bounds-Checking Interfaces
POSIX
Microsoft
Summary
Chapter 8: Input/Output
Standard I/O Streams
Stream Buffering
Predefined Streams
Stream Orientation
Text and Binary Streams
Opening and Creating Files
The fopen Function
The POSIX open Function
Closing Files
The fclose Function
The POSIX close Function
Reading and Writing Characters and Lines
Stream Flushing
Setting the Position in a File
Removing and Renaming Files
Using Temporary Files
Reading Formatted Text Streams
Reading to and Writing from Binary Streams
Summary
Chapter 9: Preprocessor
The Compilation Process
File Inclusion
Quoted and Angle Bracket Include Strings
Conditional Inclusion
Generating Errors
Using Header Guards
Macro Definitions
Macro Replacement
Type-Generic Macros
Predefined Macros
Summary
Chapter 10: Program Structure
Principles of Componentization
Coupling and Cohesion
Code Reuse
Data Abstractions
Opaque Types
Executables
Linkage
Structuring a Simple Program
Building the Code
Summary
Chapter 11: Debugging, Testing, and Analysis
Assertions
Static Assertions
Runtime Assertions
Compiler Settings and Flags
GCC and Clang
Visual C++
Debugging
Unit Testing
Static Analysis
Dynamic Analysis
AddressSanitizer
Exercises
Summary
References
Index

Citation preview

EFFECTIVE C A N

I N T R O D U C T I O N

P R O F E S S I O N A L

C

T O

P R O G R A M M I N G

ROBERT C. SE ACORD

EFFECTIVE C

EFFECTIVE C An Introduction to Professional C Programming

b y Rob e r t C . S e a c o r d

San Francisco

EFFECTIVE C. Copyright © 2020 by Robert C. Seacord. All rights reserved. No part of this work may be reproduced or transmitted in any form or by any means, electronic or mechanical, including photocopying, recording, or by any information storage or retrieval system, without the prior written permission of the copyright owner and the publisher. ISBN-10: 1-7185-0104-8 ISBN-13: 978-1-71850-104-1 Publisher: William Pollock Executive Editor: Barbara Yien Production Manager: Laurel Chun Production Editor: Katrina Taylor Cover Illustration: Gina Redman Interior Design: Octopod Studios Project Editor: Dapinder Dosanjh Developmental Editors: Liz Chadwick and Frances Saux Technical Reviewer: Martin Sebor Copyeditor: Sharon Wilkey Compositor: Happenstance Type-O-Rama Proofreader: Emelie Battaglia Indexer: Chelsea and Drew Hoffman For information on distribution, translations, or bulk sales, please contact No Starch Press, Inc. directly: No Starch Press, Inc. 245 8th Street, San Francisco, CA 94103 phone: 1.415.863.9900; [email protected] www.nostarch.com Library of Congress Cataloging-in-Publication Data Names: Seacord, Robert C., author. Title: Effective C : an introduction to professional C programming / Robert C. Seacord. Description: San Francisco : No Starch Press, Inc., 2020. | Includes bibliographical references and index. | Summary: “An in-depth guide for developers who already know how to program and want to quickly come up to speed on professional programming in C. It introduces the language’s basic data types and code structures ; addresses best practices and common errors ; and teaches how to debug, test, and analyze C programs” Provided by publisher. Identifiers: LCCN 2020017146 (print) | LCCN 2020017147 (ebook) | ISBN 9781718501041 (paperback) | ISBN 1718501048 (paperback) | ISBN 9781718501058 (ebook) Subjects: LCSH: C (Computer program language) Classification: LCC QA76.73.C15 S417 2020 (print) | LCC QA76.73.C15 (ebook) | DDC 005.13/3--dc23 LC record available at https://lccn.loc.gov/2020017146 LC ebook record available at https://lccn.loc.gov/2020017147

No Starch Press and the No Starch Press logo are registered trademarks of No Starch Press, Inc. Other product and company names mentioned herein may be the trademarks of their respective owners. Rather than use a trademark symbol with every occurrence of a trademarked name, we are using the names only in an editorial fashion and to the benefit of the trademark owner, with no intention of infringement of the trademark. The information in this book is distributed on an “As Is” basis, without warranty. While every precaution has been taken in the preparation of this work, neither the author nor No Starch Press, Inc. shall have any liability to any person or entity with respect to any loss or damage caused or alleged to be caused directly or indirectly by the information contained in it.

Dedicated to my granddaughters, Olivia and Isabella, and to other young women who will grow up to be scientists and engineers

About the Author Robert Seacord ([email protected]) is a Technical Director at NCC Group, where he develops and delivers secure coding training in C, C++, and other languages. Robert is also an expert on ISO/IEC JTC1/SC22/WG14, the international standardization working group for the C programming language. He is the author of other books, including The CERT C Coding S­tandard, Second Edition (Addison-Wesley, 2014), Secure Coding in C and C++, Second Edition (Addison-Wesley, 2013), and Java Coding Guidelines: 75 Recommendations for Reliable and Secure Programs (Addison-Wesley, 2014). He has also published more than 50 papers on software security, component-based software engineering, web-based system design, legacy-system modernization, component repositories and search engines, and user interface design and development. 

About the Contributor Aaron Ballman ([email protected]) is a compiler frontend engineer for GrammaTech, Inc. and works primarily on the static analysis tool CodeSonar. He is also a frontend maintainer for Clang, a popular open source compiler for C, C++, and other languages. Aaron is an expert for the JTC1/SC22/WG14 C programming language and JTC1/SC22/WG21 C++ programming language standards committees. His primary professional focus has been in helping programmers recognize mistakes in their code through better language design, diagnostics, and tooling. When not thinking about programming, he enjoys spending quiet moments in the woods of rural Maine with his family.

About the Tech Reviewer Martin Sebor is Principal Software Engineer with the GNU Toolchain Team at Red Hat.  His main focus is the GCC compiler and, in particular, the areas of detecting, diagnosing, preventing security-related issues in C and C++ programs, and implementing optimizations of string-based algorithms.  Prior to joining Red Hat in 2015, he worked as a compiler toolchain engineer at Cisco.  Martin has been a member of the C++ standardization committee since 1999 and a member of the C language committee since 2010.  He lives with his wife near the small town of Lyons, Colorado.

BRIEF CONTENTS

Foreword by Pascal Cuoq . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xvii Foreword by Ollie Whitehouse . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xix Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxi Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxiii Chapter 1: Getting Started with C . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 Chapter 2: Objects, Functions, and Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 Chapter 3: Arithmetic Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35 Chapter 4: Expressions and Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57 Chapter 5: Control Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81 Chapter 6: Dynamically Allocated Memory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99 Chapter 7: Characters and Strings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119 Chapter 8: Input/Output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147 Chapter 9: Preprocessor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169 Chapter 10: Program Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185 Chapter 11: Debugging, Testing, and Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223 Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227

CONTE NT S IN DE TA IL FOREWORD BY PASCAL CUOQ

XVII

FOREWORD BY OLLIE WHITEHOUSE

XIX

ACKNOWLEDGMENTS XXI INTRODUCTION XXIII A Brief History of C . . . . . . . . The C Standard . . . . . . . . . . The CERT C Coding Standard . Who This Book Is For . . . . . . What’s in This Book . . . . . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

xxiv . xxv xxvi .xxvi xxvi

Developing Your First C Program . . . . . . . . . . . . . . . . . . . . Compiling and Running Your Program . . . . . . . . . . Preprocessor Directives . . . . . . . . . . . . . . . . . . . . The main Function . . . . . . . . . . . . . . . . . . . . . . . Checking Function Return Values . . . . . . . . . . . . . Formatted Output . . . . . . . . . . . . . . . . . . . . . . . . Editors and Integrated Development Environments . . . . . . . . Compilers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . GNU Compiler Collection . . . . . . . . . . . . . . . . . . Clang . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Microsoft Visual Studio . . . . . . . . . . . . . . . . . . . . Portability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Implementation-Defined Behavior . . . . . . . . . . . . . Unspecified Behavior . . . . . . . . . . . . . . . . . . . . . Undefined Behavior . . . . . . . . . . . . . . . . . . . . . . Locale-Specific Behavior and Common Extensions . Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

1 GETTING STARTED WITH C

1

2 OBJECTS, FUNCTIONS, AND TYPES Objects, Functions, Types, and Pointers . . . . . . Declaring Variables . . . . . . . . . . . . . . . . . . . Swapping Values (First Attempt) . . . . Swapping Values (Second Attempt) . Scope . . . . . . . . . . . . . . . . . . . . . . . . . . . . Storage Duration . . . . . . . . . . . . . . . . . . . . . Alignment . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . .

. . . . . . .

. 1 . 2 . 3 . 3 . 5 . 5 . 6 . 8 . 8 . 8 . 9 . 9 10 10 10 11 11

13 . . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

13 14 15 16 18 19 20

Object Types . . . . . . . . . . Boolean Types . . . Character Types . . Numerical Types . void Types . . . . . . Function Types . . . . . . . . . Derived Types . . . . . . . . . . Pointer Types . . . . Arrays . . . . . . . . Structures . . . . . . Unions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . const . . . . . . . . . volatile . . . . . . . . restrict . . . . . . . . Exercises . . . . . . . . . . . . . Summary . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

3 ARITHMETIC TYPES

35

Integers . . . . . . . . . . . . . . . . . . . . . . . . . . Padding and Precision . . . . . . . . . The Header File . . . . . . . Declaring Integers . . . . . . . . . . . . Unsigned Integers . . . . . . . . . . . . Signed Integers . . . . . . . . . . . . . . Integer Constants . . . . . . . . . . . . . Floating-Point . . . . . . . . . . . . . . . . . . . . . . Floating-Point Types . . . . . . . . . . . Floating-Point Arithmetic . . . . . . . . Floating-Point Values . . . . . . . . . . . Floating-Point Constants . . . . . . . . Arithmetic Conversion . . . . . . . . . . . . . . . . Integer Conversion Rank . . . . . . . . Integer Promotions . . . . . . . . . . . . Usual Arithmetic Conversions . . . . . An Example of Implicit Conversion . Safe Conversions . . . . . . . . . . . . . Summary . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

4 EXPRESSIONS AND OPERATORS Simple Assignment . . . . . . . . . . . . . . . Evaluations . . . . . . . . . . . . . . . . . . . . . Function Invocation . . . . . . . . . . . . . . . Increment and Decrement Operators . . . Operator Precedence and Associativity .

xii   Contents in Detail

21 21 22 22 24 24 25 25 26 28 29 30 32 32 33 33 34 34

. . . . .

. . . . .

. . . . .

. . . . .

36 36 36 37 37 40 44 45 46 47 47 49 49 50 51 52 53 54 55

57 . . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

58 59 60 61 62

Order of Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Unsequenced and Indeterminately Sequenced Evaluations . Sequence Points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . sizeof Operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Arithmetic Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Unary + and – Operators . . . . . . . . . . . . . . . . . . . . . . . . Logical Negation Operator . . . . . . . . . . . . . . . . . . . . . . . Multiplicative Operators . . . . . . . . . . . . . . . . . . . . . . . . . Additive Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Bitwise Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Complement Operator . . . . . . . . . . . . . . . . . . . . . . . . . . Shift Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Bitwise AND Operator . . . . . . . . . . . . . . . . . . . . . . . . . . Bitwise Exclusive OR Operator . . . . . . . . . . . . . . . . . . . . . Bitwise Inclusive OR Operator . . . . . . . . . . . . . . . . . . . . . Logical Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Cast Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Conditional Operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . _Alignof Operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Relational Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Compound Assignment Operators . . . . . . . . . . . . . . . . . . . . . . . . . Comma Operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Pointer Arithmetic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

5 CONTROL FLOW Expression Statements . . . . . . . . . Compound Statements . . . . . . . . . Selection Statements . . . . . . . . . . The if Statement . . . . . . . The switch Statement . . . Iteration Statements . . . . . . . . . . . The while Statement . . . . The do...while Statement . The for Statement . . . . . . Jump Statements . . . . . . . . . . . . . The goto Statement . . . . . The continue Statement . . The break Statement . . . . The return Statement . . . . Exercises . . . . . . . . . . . . . . . . . . Summary . . . . . . . . . . . . . . . . . .

64 65 66 66 67 67 68 68 69 69 70 70 72 72 73 73 75 76 76 77 78 78 79 80

81 . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

6 DYNAMICALLY ALLOCATED MEMORY

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

81 82 83 83 86 89 89 90 91 93 93 94 95 96 97 97

99

Storage Duration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100 The Heap and Memory Managers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100 When to Use Dynamically Allocated Memory . . . . . . . . . . . . . . . . . . . . . . . 101

Contents in Detail   xiii

Memory Management Functions . . . . . . The malloc Function . . . . . . . . The aligned_alloc Function . . . The calloc Function . . . . . . . . . The realloc Function . . . . . . . . The reallocarray Function . . . . The free Function . . . . . . . . . . Memory States . . . . . . . . . . . . . . . . . . Flexible Array Members . . . . . . . . . . . . Other Dynamically Allocated Storage . . The alloca Function . . . . . . . . Variable-Length Arrays . . . . . . Debugging Allocated Storage Problems . Dmalloc . . . . . . . . . . . . . . . . Safety-Critical Systems . . . . . . Exercises . . . . . . . . . . . . . . . . . . . . . . Summary . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . .

Characters . . . . . . . . . . . . . . . . . . . . . . . . . . ASCII . . . . . . . . . . . . . . . . . . . . . . . Unicode . . . . . . . . . . . . . . . . . . . . . Source and Execution Character Sets . Data Types . . . . . . . . . . . . . . . . . . . Character Constants . . . . . . . . . . . . . Escape Sequences . . . . . . . . . . . . . . Linux . . . . . . . . . . . . . . . . . . . . . . . Windows . . . . . . . . . . . . . . . . . . . . Character Conversion . . . . . . . . . . . . Strings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . String Literals . . . . . . . . . . . . . . . . . . String-Handling Functions . . . . . . . . . . . . . . . . and . . . . . . . . Annex K Bounds-Checking Interfaces . POSIX . . . . . . . . . . . . . . . . . . . . . . . Microsoft . . . . . . . . . . . . . . . . . . . . . Summary . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . .

7 CHARACTERS AND STRINGS

119

8 INPUT/OUTPUT Standard I/O Streams . . . . . . . . . . Stream Buffering . . . . . . . Predefined Streams . . . . . . Stream Orientation . . . . . . Text and Binary Streams . . Opening and Creating Files . . . . . . The fopen Function . . . . . . The POSIX open Function .

xiv   Contents in Detail

101 102 104 105 105 107 108 109 110 111 111 112 115 116 118 118 118

120 120 120 122 122 124 125 126 126 128 131 132 134 .135 141 144 145 145

147 . . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

148 148 149 150 150 151 151 153

Closing Files . . . . . . . . . . . . . . . . . . . . . . . . The fclose Function . . . . . . . . . . . . . The POSIX close Function . . . . . . . . Reading and Writing Characters and Lines . . . Stream Flushing . . . . . . . . . . . . . . . . . . . . . Setting the Position in a File . . . . . . . . . . . . . Removing and Renaming Files . . . . . . . . . . . Using Temporary Files . . . . . . . . . . . . . . . . . Reading Formatted Text Streams . . . . . . . . . . Reading to and Writing from Binary Streams . Summary . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

9 PREPROCESSOR

169

The Compilation Process . . . . . . . . . . . . . . . . . . . . File Inclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Quoted and Angle Bracket Include Strings . Conditional Inclusion . . . . . . . . . . . . . . . . . . . . . . . Generating Errors . . . . . . . . . . . . . . . . . . . Using Header Guards . . . . . . . . . . . . . . . . Macro Definitions . . . . . . . . . . . . . . . . . . . . . . . . . Macro Replacement . . . . . . . . . . . . . . . . . Type-Generic Macros . . . . . . . . . . . . . . . . Predefined Macros . . . . . . . . . . . . . . . . . . Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

10 PROGRAM STRUCTURE . . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

11 DEBUGGING, TESTING, AND ANALYSIS . . . . . . . .

. . . . . . . .

. . . . . . . .

170 171 172 172 173 174 175 178 180 181 183

185

Principles of Componentization . . . Coupling and Cohesion . Code Reuse . . . . . . . . . . Data Abstractions . . . . . . Opaque Types . . . . . . . . Executables . . . . . . . . . . . . . . . . Linkage . . . . . . . . . . . . . . . . . . . Structuring a Simple Program . . . . Building the Code . . . . . . . . . . . . Summary . . . . . . . . . . . . . . . . . .

Assertions . . . . . . . . . . . . . Static Assertions . . . Runtime Assertions . Compiler Settings and Flags . GCC and Clang . . Visual C++ . . . . . . Debugging . . . . . . . . . . . . . Unit Testing . . . . . . . . . . . . .

154 154 155 155 157 158 161 161 162 165 168

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

186 186 187 187 188 189 191 192 196 198

199 199 200 202 203 204 206 208 211

Contents in Detail   xv

Static Analysis . . . . . . . . . Dynamic Analysis . . . . . . . AddressSanitizer . Exercises . . . . . . . . . . . . . Summary . . . . . . . . . . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

214 216 217 221 221

REFERENCES

223

INDEX

227

xvi   Contents in Detail

FORE WORD BY PA SC A L CUOQ

The first time I came across Robert Seacord’s name was in 2008. Robert was already well known in the C programming galaxy for his work on the CERT C Coding Standard and Annex K of the C Standard. But in 2008, it had been only a few years since—young and foolish—I had embarked on the Frama-C p ­ roject to guarantee the absence of undefined behavior in C programs. At some point, a CERT Vulnerability Note about how C compilers (and GCC in particular) removed certain pointer-arithmetic overflow checks, piqued my interest. The compilers had reason to eliminate the checks: naively written, they invoked undefined behavior when the overflow was present.

The C compilers were also allowed to tell the programmer nothing about what they had done wrong, even at the maximum warning level. Undefined behavior in C can be harsh. I had set out to solve this exact problem. Robert was one of the authors of that note. Effective C will teach you C programming for the modern era. It will help you establish good habits to keep you from using undefined behavior, whether voluntarily or through negligence. Let the reader be warned: in large C programs, avoiding ordinary programming errors alone may not suffice for dodging undefined behavior caused by arbitrary inputs! This book’s emphasis on the security aspects of C programming is unmatched. My personal recommendation is that, after reading it, you use all of the available tools it presents to avoid undefined behavior in the C programs you write. Pascal Cuoq Chief Scientist, TrustInSoft

xviii   Foreword by Pascal Cuoq

FORE WORD BY OLLIE W HITEHOUSE

When I started in cyber security over 25 years ago, I learned my trade primarily by finding and exploiting unsafe memory handling in C programs—a class of vulnerability that, even at the time, was over 20 years old. In my career at BlackBerry, as I waded through torrents of code review, I saw firsthand how dangerous C could be to the improperly initiated. Now, as a CTO for a multinational cyber security professional and managed services firm, I see the consequences of poorly written C code on our connected society every day. Today we still face numerous challenges in writing secure and professional C. The many innovations in compiler- and operating system–level mitigations can be and are regularly undermined. And even while we see

advanced innovations in other modern languages, there is still a growing demand for C, particularly in IoT or other highly resource constrained environments. Robert is the authority on how to program professionally and securely in C. For over a decade, I have recommended his material to customers and internal teams alike. There is no better person to teach how to code C in a professional and, among other things, secure manner. Writing professional C today means writing code that is performant, safe, and secure. By doing so, you will be able to contribute to our connected society without increasing its technical debt. This book will help those with little or no C experience quickly develop the knowledge and skills to become professional C programmers and will provide a strong foundation for developing systems that are performant, safe, and secure. Ollie Whitehouse Global CTO, NCC Group

xx   Foreword by Ollie Whitehouse

ACKNOWLEDGMENTS

I would like to acknowledge the contributions of all the folks who made this possible. I’ll start with Bill Pollock at No Starch Press, who doggedly pursued me to write a C book. I would like to thank Ollie Whitehouse and Pascal Cuoq for the excellent forewords. Aaron Ballman has been a valuable partner for the duration of this effort. In addition to contributing two chapters, he reviewed everything (often multiple times) and helped me solve problems from the profound to the mundane. Douglas Gwyn, Emeritus at the US Army Research Laboratory and also an Emeritus member of the C Standards Committee, helped review all the chapters. When my writing was not up to his standards, he guided me in the right direction. Martin Sebor is a good friend who served as the official technical reviewer for the book, so any inaccuracies you may find are definitely his fault. In addition to Aaron, Doug, and Martin, several other distinguished members of the C and C++ Standards Committee reviewed chapters including Jim Thomas, Thomas Köppe, Niall Douglas, Tom Honermann, and JeanHeyd Meneide. Technical reviewers among my colleagues at NCC Group include Nick Dunn, Jonathan Lindsay, Tomasz Kramkowski, Alex

Donisthorpe, Joshua Dow, Catalin Visinescu, Aaron Adams, and Simon Harraghy. Technical reviewers not associated with any of these organizations include David LeBlanc, Nicholas Winter, John McFarlane, and Scott Aloisio. I would also like to thank the following professionals at No Starch who ensured a quality product: Elizabeth Chadwick, Frances Saux, Zach Lebowski, Annie Choi, Barbara Yien, Katrina Taylor, Natalie Gleason, Derek Yee, Laurel Chun, Gina Redman, Sharon Wilkey, Emelie Battaglia, and Dapinder Dosanjh. Finally, I would like to thank Drew and Chelsea Hoffman for assisting with the indexing.

xxii   Acknowledgments

INTRODUCTION

C was developed as a system programming language in the 1970s, and even after all this time, it remains incredibly popular. System languages are designed for performance and ease of access to the underlying hardware while providing high-level programming features. While other languages may offer newer language features, their compilers and libraries are typically written in C. Carl Sagan once said, “If you wish to make an apple pie from scratch, you must first invent the universe.” The inventors of C did not invent the universe; they designed C to work with a variety of computing hardware and architectures that, in turn, were constrained by physics and mathematics. C is layered directly on top of computing hardware, making it more sensitive to evolving hardware features, such as vectorized instructions, than higher-level languages that typically rely on C for their efficiency.

According to the TIOBE index, C has been either the most popular pro­ gramming language or second most popular since 2001.1 C is TIOBE’s programming language of the year for 2019. The popularity of the C programming language can most likely be attributed to several tenets of the language referred to as the spirit of C: •

Trust the programmer. Generally speaking, the C language assumes you know what you’re doing and lets you. This isn’t always a good thing (for example, if you don’t know what you’re doing).



Don’t prevent the programmer from doing what needs to be done. Because C is a system programming language, it has to be able to handle a variety of low-level tasks. Keep the language small and simple. The language is designed to be fairly close to the hardware and to have a small footprint. Provide only one way to do an operation. Also known as conservation of mechanism, the C language tries to limit the introduction of duplicate mechanisms. Make it fast, even if it isn’t guaranteed to be portable. Allowing you to write optimally efficient code is the top priority. The responsibility of ensuring that code is portable, safe, and secure is delegated to you, the programmer.

• •



A Brief History of C The C programming language was developed in 1972 by Dennis Ritchie and Ken Thompson at Bell Telephone Laboratories. Brian Kernighan coauthored The C Programming Language (K&R 1988) with Dennis Ritchie. In 1983, the American National Standards Institute (ANSI) formed the X3J11 committee to establish a standard C specification, and in 1989, the C Standard was ratified as ANSI X3.159-1989, “Programming Language C.” This 1989 version of the language is referred to as ANSI C or C89. In 1990, the ANSI C Standard was adopted (unchanged) by a joint technical committee of the International Organization for Standardization (ISO) and the International Electrotechnical Commission (IEC) and published as the first edition of the C Standard, C90 (ISO/IEC 9899:1990). The second edition of the C Standard, C99, was published in 1999 (ISO/IEC 9899:1999), and a third edition, C11, in 2011 (ISO/IEC 9899:2011). The latest version of the C Standard (as of this writing) is the fourth version, published in 2018 as C17 (ISO/IEC 9899:2018). A new major revision referred to as C2x is under

1. T  he TIOBE Programming Community index is an indicator of the popularity of programming languages and can be found at https://www.tiobe.com/tiobe-index/. The ratings are a measure of the number of skilled engineers, courses, and third party vendors for each language. The index can be used to help you decide which programming languages to learn or adopt when building a new software system.

xxiv   Introduction

development by ISO/IEC. According to 2018 polling data from JetBrains, 52 percent of C programmers use C99, 36 percent use C11, and 23 percent use an embedded version of C.2

The C Standard The C Standard (ISO/IEC 9899:2018) defines the language and is the final authority on language behavior. While the standard can be obscure to impenetrable, you need to understand it if you intend to write code that’s portable, safe, and secure. The C Standard provides a substantial degree of latitude to implementations to allow them to be optimally efficient on various hardware platforms. Implementations is the term used by the C Standard to refer to compilers and is defined as follows: A particular set of software, running in a particular translation environment under particular control options, that performs translation of programs for, and supports execution of functions in, a particular execution environment.

This definition indicates that each compiler with a particular set of command line flags, along with the C Standard Library, is considered a separate implementation, and different implementations can have significantly different implementation-defined behavior. This is noticeable in GNU Compiler Collection (GCC), which uses the -std= flag to determine the language standard. Possible values for this option include c89, c90, c99, c11, c17, c18, and c2x. The default depends on the version of the compiler. If no C language dialect options are given, the default for GCC 10 is -std=gnu17, which provides extensions to the C language. For portability, specify the standard you’re using. For access to new language features, specify a recent standard. A good choice (in 2019) with GCC 8 and later is -std=c17. Because implementations have such a range of behaviors, and because some of these behaviors are undefined, you can’t understand the C language by just writing simple test programs to examine the behavior.3 The behavior of the code may vary when compiled by a different implementation on different platforms or even the same implementation using a different set of flags or a different C Standard Library implementation. Code behavior can even vary between versions of a compiler. The C Standard is the only document that specifies which behaviors are guaranteed for all implementations, and where you need to plan for variability. This is mostly a concern when developing portable code, but can also affect the security and safety of your code.

2. S ee the JetBrains website for more details on the survey: https://www.jetbrains.com/lp/devecosystem-2019/c/. 3. If you want to try this, the Compiler Explorer is an excellent tool; see https://godbolt.org/. Introduction   xxv

The CERT C Coding Standard The CERT ® C Coding Standard, Second Edition: 98 Rules for Developing Safe, Reliable, and Secure Systems (Seacord 2014) is a reference book I wrote while managing the secure coding team at the Software Engineering Institute at Carnegie Mellon University. The CERT C Coding Standard contains examples of common C programming mistakes and how to correct them. Throughout this book, we reference some of these rules as a source for detailed information on specific C language programming topics.

Who This Book Is For This book is an introduction to the C language. It is written to be as accessible as possible to anyone who wants to learn C programming, without dumbing it down. In other words, we didn’t overly simplify C programming in the way many other introductory books and courses might. These overly simplified references will teach you how to get code to compile and run, but the code might still be wrong. Developers who learn how to program C from such sources will typically develop substandard, flawed, insecure code that will eventually need to be rewritten (often sooner than later). Hopefully, these developers will eventually benefit from senior developers in their organizations who will help them unlearn these harmful misconceptions about programming in C, and help them start developing professional quality C code. On the other hand, this book will quickly teach you how to develop correct, portable, professional-quality code, build a foundation for developing security-critical and safety-critical systems, and perhaps teach you a thing or two that even the senior developers at your organization don’t know. Effective C: An Introduction to Professional C Programming is a concise introduction to essential C language programming that will soon have you writing programs, solving problems, and building working systems. The code examples are idiomatic and straightforward. In this book, you’ll learn about essential programming concepts in C and practice writing high-quality code with exercises for each topic. You’ll also learn about good software engineering practices for developing correct, secure C code. Go to this book’s page at https://www.nostarch.com /effective_c/ or to http://www.robertseacord.com/ where we will provide updates and additional material. If you are interested in learning more about secure coding in C, C++, or other languages after completing this book, please check out the training classes offered through NCC Group at https://www .nccgroup.trust/us/our-services/cyber-security/security-training/secure-coding/.

What’s in This Book This book starts with an introductory chapter that covers just enough material to get you programming right from the start. After this, we circle back

xxvi   Introduction

and examine the basic building blocks of the language. The book culminates with two chapters that will show you how to compose real-world systems from these basic building blocks and how to debug, test, and analyze the code you’ve written. The chapters are as follows: Chapter 1: Getting Started with C  You’ll write a simple C program to become familiar with using the main function. You’ll also look at a few options for editors and compilers. Chapter 2: Objects, Functions, and Types  This chapter explores basics like declaring variables and functions. You’ll also look into the principles of using basic types. Chapter 3: Arithmetic Types  You’ll learn about the two kinds of arithmetic data types: integers and floating-point types. Chapter 4: Expressions and Operators  You’ll learn about operators and how to write simple expressions to perform operations on various object types. Chapter 5: Control Flow  You’ll learn how to control the order in which individual statements are evaluated. We’ll start by going over expression statements and compound statements that define the work to be performed. We’ll then cover three kinds of statements that determine which code blocks are executed, and in what order: selection, iteration, and jump statements. Chapter 6: Dynamically Allocated Memory  You’ll learn about dynamically allocated memory, which is allocated from the heap at runtime. Dynamically allocated memory is useful when the exact storage requirements for a program are unknown before runtime. Chapter 7: Character and Strings  You’ll learn about the various character sets, including ASCII and Unicode, that can be used to compose strings. You’ll learn how strings are represented and manipulated using the legacy functions from the C Standard Library, the bounds-checking interfaces, and POSIX and Windows APIs. Chapter 8: Input/Output  This chapter will teach you how to perform input/output (I/O) operations to read data from, or write data to, terminals and filesystems. I/O involves all the ways information enters or exits a program, without which your programs would be useless. We’ll cover techniques that make use of C Standard streams and POSIX file descriptors. Chapter 9: Preprocessor  You’ll learn how to use the preprocessor to include files, define object- and function-like macros, and conditionally include code based on implementation-specific features. Chapter 10: Program Structure  You’ll learn how to structure your program into multiple translation units consisting of both source and include files. You’ll also learn how to link multiple object files together to create libraries and executable files.

Introduction   xxvii

Chapter 11: Debugging, Testing, and Analysis  This chapter describes tools and techniques for producing correct programs, including c­ompiletime and runtime assertions, debugging, testing, static analysis, and dynamic analysis. The chapter also discusses which compiler flags are recommended for use in different phases of the software development process. You’re about to embark on a journey from which you will emerge a newly minted but professional C developer.

xxviii   Introduction

1 GE T TING S TA R TE D W ITH C

In this chapter, you’ll develop your first C program: the traditional “Hello, world!” program. I’ll take you through the various aspects of this simple program, as well as compiling and running a C program. Then I’ll discuss some editor and compiler options, and lay out common portability issues you’ll quickly become familiar with as you code in C. Developing Your First C Program The best way to learn C programming is to start writing C programs, and the traditional program to start with is “Hello, world!” To write this program, you need a text editor or integrated development environment (IDE). There are many to choose from, but for now, open your favorite editor, and we’ll survey other options later in this chapter.

In your text editor, enter the program in Listing 1-1. #include #include u int main(void) { v puts("Hello, world!"); w return EXIT_SUCCESS; x } Listing 1-1: The hello.c program

We’ll go over each line of this program in more detail shortly. For now, save this file as hello.c. The file extension .c indicates that the file contains C language source code. NOTE

If you’ve purchased an ebook, cut and paste the program into the editor. Use cut and paste when possible, as it can cut down on transcription errors.

Compiling and Running Your Program Next we need to compile and run the program, which involves two separate steps. You can choose from numerous C compilers, and the command to compile the program depends on which compiler you’re using. On Linux and other Unix-like operating systems, you can invoke the system compiler with the cc command. To compile your program, enter cc on the command line followed by the name of the file you want to compile: % cc hello.c

NOTE

These commands are specific to Linux (and other Unix-like operating systems). Other compilers on other operating systems will need to be invoked differently. Refer to the documentation for your specific compiler. If you entered the program correctly, the compile command will create a new file called a.out in the same directory as your source code. Inspect your directory by using the ls command and you should see the following: % ls a.out hello.c

The a.out file is the executable program, which you can now run on the command line: % ./a.out Hello, world!

2   Chapter 1

If everything goes right, the program should print Hello, world! to the terminal window. If it doesn’t, compare the program text from Listing 1-1 to your program to ensure they are the same. The cc command has numerous flags and compiler options. The -o file flag, for example, lets you give the executable file a memorable name instead of a.out. The following compiler invocation names the executable hello: % cc -o hello hello.c % ./hello Hello, world!

Now we’ll inspect the hello.c program line by line.

Preprocessor Directives The first two lines of the hello.c program use the #include preprocessor directive, which behaves as if you replaced it with the contents of the specified file at the exact same location. We include the and headers to access the functions declared in those headers, which we can then call from the program. The puts function is declared in , and the EXIT_SUCCESS macro is defined in . As the filenames suggest, contains the declarations for C Standard I/O functions, and contains the declarations for general utility functions. You need to include the declarations for any library functions that you use in your program.

The main Function The main portion of the program, shown previously in Listing 1-1, starts with u: int main(void) {

This line defines the main function that’s called at program startup. The main function defines the main entry point for the program that’s executed in a hosted environment when the program is invoked from the command line or from another program. C defines two possible execution environments: freestanding and hosted. A freestanding environment may not provide an operating system and is typically used in embedded programming. These implementations provide a minimal set of library functions, and the name and type of the function called at program startup are implementation defined. This book primarily assumes a hosted environment. We define main to return a value of type int and place void inside the parenthesis to indicate that the function does not accept arguments. The int type is a signed integer type that can be used to represent both positive and negative integer values as well as zero. Similar to other procedural languages, C programs consist of procedures (referred to as functions) that can accept arguments and return values. Each function is a reusable unit of work that you can invoke as frequently as necessary in your program. In this case, the

Getting Started with C   3

value returned by the main function indicates whether the program terminated successfully. The actual work performed by this particular function v is to print out the line Hello, world!: puts("Hello, world!");

The puts function is a C Standard Library function that writes a string argument to stdout, which typically represents the console or terminal window, and appends a newline character to the output. "Hello, world!" is a string literal that behaves like a read-only string. This function invocation outputs Hello, world! to the terminal. Once your program has completed, you’ll want it to exit. You can exit the program using the return statement w within main to return an integer value to the host environment or invoking script: return EXIT_SUCCESS; EXIT_SUCCESS is an object-like macro that commonly expands to 0 and is typically defined as follows: #define EXIT_SUCCESS 0

Each occurrence of EXIT_SUCCESS is replaced by a 0, which is then returned to the host environment from the call to main. The script that invokes the program can then check its status to determine whether the invocation was successful. A return from the initial call to the main function is equivalent to calling the C Standard Library exit function with the value returned by the main function as its argument. The final line of this program x consists of a closing brace }, which closes the code block we started with the declaration of the main function: int main(void) { // ---snip--}

You can place the opening brace on the same line as the declaration or its own line, as follows: int main(void) { // ---snip--}

This decision is strictly a stylistic one, because whitespace characters (including newlines) are generally not syntactically meaningful. In this book, I generally place the opening brace on the line with the function ­declaration because it’s stylistically more compact.

4   Chapter 1

Checking Function Return Values Functions will often return a value that’s the result of a computation or that signifies whether the function successfully completed its task. For example, the puts function we used in our “Hello, world!” program takes a string to print and returns a value of type int. The puts function returns the value of the macro EOF (a negative integer) if a write error occurs; otherwise, it returns a nonnegative integer value. Although it’s unlikely that the puts function will fail and return EOF for our simple program, it’s possible. Because the call to puts can fail and return EOF, it means that your first C program has a bug or, at least, can be improved as follows: #include #include int main(void) { if (puts("Hello, world!") == EOF) { return EXIT_FAILURE; // code here never executes } return EXIT_SUCCESS; // code here never executes }

This revised version of the “Hello, world!” program checks whether the puts call returns the value EOF, indicating a write error. If the function returns EOF, the program returns the value of the EXIT_FAILURE macro (which evaluates to a nonzero value). Otherwise, the function succeeds, and the program returns EXIT_SUCCESS (which is required to be 0). The script that invokes the program can then check its status to determine whether it was successful. Code following a return statement is dead code that never executes. This is indicated by single line comments in the revised program. Everything following // is ignored by the compiler.

Formatted Output The puts function is a nice, simple way to write a string to stdout, but eventually you’ll need to print formatted output by using the printf function—for example, to print arguments other than strings. The printf function takes a format string that defines how the output is formatted, followed by a variable number of arguments that are the actual values you want to print. For example, if you want to use the printf function to print out Hello, world!, you could write it like this: printf("%s\n", "Hello, world!");

The first argument is the format string "%s\n". The %s is a conversion specification that instructs the printf function to read the second argument (a string literal) and print it to stdout. The \n is an alphabetic escape sequence used to represent nongraphic characters, and tells the function to Getting Started with C   5

include a new line after the string. Without the newline sequence, the next characters printed (likely the command prompt) would appear on the same line. This function call outputs the following: Hello, world!

Take care not to pass user-supplied data as part of the first argument to the printf function, because doing so can result in a formatted output security vulnerability (Seacord 2013). The simplest way to output a string is to use the puts function, as previously shown. If you do use printf instead of puts in the revised version of the “Hello, world!” program, you’ll find it no longer works, because the printf function returns status differently than the puts function. The printf function returns the number of characters printed if it’s successful, or a negative value if an output or encoding error occurred. You can try modifying the “Hello, world!” program to use the printf function as an exercise.

Editors and Integrated Development Environments A variety of editors and IDEs can be used to develop your C programs. Figure 1-1 shows the most used editors, according to a 2018 JetBrains survey. Which IDE/editor do you use the most? Other vi/Vim

Emacs Xcode Sublime Text Atom

Visual Studio Code CLion Visual Studio

Figure 1-1: IDE/editor use

The exact tools available depend on which system you’re using. This book focuses on Linux, Windows, and macOS, as these are the most common development platforms. For Microsoft Windows, Microsoft’s Visual Studio IDE (https://visualstudio .microsoft.com/) is an obvious choice. Visual Studio comes in three editions: Community, Professional, and Enterprise. The Community edition has the advantage of being free, while the other editions add features at a cost. For this book, you’ll need only the Community edition.

6   Chapter 1

For Linux, the choice is less obvious. Vim, Emacs, Visual Studio Code, and Eclipse are all options. Vim is the editor of choice for many developers and power users. It is a text editor based on the vi editor written by Bill Joy in the 1970s for a version of Unix. It inherits the key bindings of vi but also adds functionality and extensibility that are missing from the original vi. You can optionally install Vim plug-ins such as YouCompleteMe (https://github.com/Valloric/YouCompleteMe/) or deoplete (https://github.com /Shougo/deoplete.nvim/) that provide native semantic code completion for C programming. GNU Emacs is an extensible, customizable, and free text editor. At its core, it’s an interpreter for Emacs Lisp, a dialect of the Lisp programming language with extensions to support text editing—although I’ve never found this to be a problem. Full disclosure: almost all the C code I’ve ever developed was edited in Emacs. Visual Studio Code (VS Code) is a streamlined code editor with support for development operations such as debugging, task running, and version control (covered in Chapter 11). It provides just the tools a developer needs for a quick code-build-debug cycle. VS Code runs on macOS, Linux, and Windows and is free for private or commercial use. Installation instructions are available for Linux and other platforms;4 on Windows, you’d most likely use Microsoft’s Visual Studio. Figure 1-2 shows Visual Studio Code being used to develop the “Hello, world!” program from Listing 1-1 on Ubuntu. You can see from the debug console that the program exited with status code 0 as expected.

Figure 1-2: Visual Studio Code running on Ubuntu

4. S ee “Visual Studio Code on Linux” at https://code.visualstudio.com/docs/setup/linux/ for installation instructions on Linux. Getting Started with C   7

Compilers Many C compilers are available, so I won’t discuss them all here. Different compilers implement different versions of the C Standard. Many compilers for embedded systems support only C89/C90. Popular compilers for Linux and Windows work harder to support modern versions of the C Standard, up to and including support for C2x.

GNU Compiler Collection The GNU Compiler Collection (GCC) includes frontends for C, C++, and Objective-C, as well as other languages (https://gcc.gnu.org/). GCC development follows a well-defined development plan under the guidance of the GCC steering committee. GCC has been adopted as the standard compiler for Linux systems, although versions are also available for Microsoft Windows, macOS, and other platforms. Installing GCC on Linux is easy. The following command, for example, should install GCC 8 on Ubuntu: % sudo apt-get install gcc-8

You can test the version of GCC you’re using with the following command: % gcc --version gcc (Ubuntu 8.3.0-6ubuntu1~18.04) 8.3.0 This is free software; see the source for copying conditions. There is NO Warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

Fedora is an ideal development system if you intend to develop software for Red Hat Enterprise Linux. The following command can be used to install GCC on Fedora: % sudo dnf install gcc

Clang Another popular compiler is Clang (https://clang.llvm.org/). Installing Clang on Linux is also easy. The following command, for example, should install Clang on Ubuntu: % sudo apt-get install clang

You can test the version of Clang you’re using with the following command: % clang --version clang version 6.0.0-1ubuntu2 (tags/RELEASE_600/final) Target: x86_64-pc-linux-gnu Thread model: posix InstalledDir: /usr/bin

8   Chapter 1

Microsoft Visual Studio The most popular development environment for Windows is Microsoft Visual Studio, which includes both the IDE and the compiler. The most recent version of Visual Studio at the time of writing is 2019 (https://visualstudio .microsoft.com/downloads/). It’s bundled with Visual C++ 2019, which includes both the C and C++ compilers. You can set options for Visual Studio on the Project Property Pages. On the Advanced tab under C/C++, make sure you compile as C code by using the Compile as C Code (/TC) option and not the Compile as C++ Code (/TP) option. By default, when you name a file with a .c extension, it’s compiled with /TC. If the file is named with a .cpp, .cxx, or a few other extensions, it’s compiled with /TP.

Portability Every C compiler implementation is at least a little different. Compilers continually evolve—so, for example, a compiler like GCC might provide full support for C17 but be working toward support for C2x, in which case it might have some C2x features implemented but not others. Consequently, compilers support a full spectrum of C Standard versions (including inbetween versions). The overall evolution of C implementations is slow, with many compilers significantly lagging behind the C Standard. Programs written for C can be considered strictly conforming if they use only those features of the language and library specified in the standard. These programs are intended to be maximally portable. However, because of the range of implementation behaviors, no real-world C program is strictly conforming, nor ever will be (and probably shouldn’t be). Instead, the C Standard allows you to write conforming programs that may depend on nonportable language and library features. It’s common practice to write code for a single reference implementation, or sometimes several implementations, depending on which platforms you plan to deploy your code. The C Standard is what ensures that these implementations don’t differ too much, and allows you to target several at once, without having to learn a new language each time. Five kinds of portability issues are enumerated in Annex J of the C Standard documents: • • • • •

Implementation-defined behavior Unspecified behavior Undefined behavior Locale-specific behavior Common extensions

As you learn about the C language, you’ll encounter examples of all five kinds of behaviors, so it’s important to understand precisely what these are.

Getting Started with C   9

Implementation-Defined Behavior Implementation-defined behavior is program behavior that’s not specified by the C Standard and that may offer different results among implementations, but has consistent, documented behavior within an implementation. An example of implementation-defined behavior is the number of bits in a byte. Implementation-defined behaviors are mostly harmless but can cause defects when porting to different implementations. Whenever possible, avoid writing code that depends on implementation-defined behaviors that vary among the C implementations you might use to compile your code. A complete list of implementation-defined behaviors is enumerated in Annex J.3 of the C Standard. You can document your dependencies on these implementation-defined behaviors by using a static_assert declaration, as discussed in Chapter 11. Throughout this book, I note when code has implementation-defined behavior.

Unspecified Behavior Unspecified behavior is program behavior for which the standard provides two or more options. The standard imposes no requirements on which option is chosen in any instance. Each execution of a given expression may have different results or produce a different value than a previous execution of the same expression. An example of unspecified behavior is function parameter storage layout, which can vary among function invocations within the same program. Avoid writing code that depends on the unspecified behaviors enumerated in Annex J.1 of the C Standard.

Undefined Behavior Undefined behavior is behavior that isn’t defined by the C Standard, or less circularly, “behavior, upon use of a nonportable or erroneous program construct or of erroneous data, for which the standard imposes no requirements.” Examples of undefined behavior include signed integer overflow and dereferencing an invalid pointer value. Code that has undefined behavior is often erroneous, but is more nuanced than that. Undefined behaviors are identified in the standard as follows: • • •

When a “shall” or “shall not” requirement is violated, and that requirement appears outside a constraint, the behavior is undefined When behavior is explicitly specified by the words “undefined behavior” By the omission of any explicit definition of behavior

The first two kinds of undefined behavior are frequently referred to as explicit undefined behaviors, while the third kind is referred to as implicit

10   Chapter 1

undefined behavior. There is no difference in emphasis among these three; they all describe behavior that is undefined. The C Standard Annex J.2, “Undefined behavior,” contains a list of explicit undefined behaviors in C. Developers frequently misunderstand undefined behaviors to be errors or omissions in the C Standard, but the decision to classify a behavior as undefined is intentional and considered. Behaviors are classified as undefined by the C Standards committee to do the following: • • •

Give the implementer license not to catch program errors that are difficult to diagnose Avoid defining obscure corner cases that would favor one implementation strategy over another Identify areas of possible conforming language extension in which the implementer may augment the language by providing a definition of the officially undefined behavior

These three reasons are really quite different but are all considered portability issues. We’ll examine examples of all three as they come up over the course of this book. Compilers (implementations) have the latitude to do the following: • • •

Ignore undefined behavior completely, giving unpredictable results Behave in a documented manner characteristic of the environment (with or without issuing a diagnostic) Terminate a translation or execution (with issuing a diagnostic)

None of these options are great (particularly the first), so it’s best to avoid undefined behaviors except when the implementation specifies these behaviors are defined to allow you to invoke a language augmentation.5

Locale-Specific Behavior and Common Extensions Locale-specific behavior depends on local conventions of nationality, culture, and language that each implementation documents. Common extensions are widely used in many systems but are not portable to all implementations.

Summary In this chapter, you learned how to write a simple C language program, compile it, and run it. We then looked at several editors and interactive development environments as well as a few compilers that you can use to develop C programs on Windows, Linux, and macOS systems. Generally speaking, you should use newer versions of the compilers and other tools,

5. C ompilers sometimes have a pedantic mode that can help notify the programmer of these portability issues. Getting Started with C   11

as they tend to support newer features of the C programming language and provide better diagnostics and optimizations. You may not want to use newer versions of compilers if they break your existing code or if you’re getting ready to deploy your code, to avoid introducing unnecessary changes into your already tested application. We concluded this chapter with a discussion of the portability of C language programs. Subsequent chapters will examine specific features of the C language and library, starting with objects, functions, and types in the next chapter.

12   Chapter 1

2 OBJEC T S, F UNC T IONS, AND T YPES

In this chapter, you’ll learn about objects, functions, and types. We’ll examine how to declare variables (objects with identifiers) and functions, take the addresses of objects, and dereference those object pointers. You’ve already seen some types that are available to C programmers. The first thing you’ll learn in this chapter is one of the last things that I learned: every type in C is either an object type or a function type. Objects, Functions, Types, and Pointers An object is storage in which you can represent values. To be precise, an object is defined by the C Standard (ISO/IEC 9899:2018) as a “region of

data storage in the execution environment, the contents of which can represent values,” with the added note, “when referenced, an object can be interpreted as having a particular type.” A variable is an example of an object. Variables have a declared type that tells you the kind of object its value represents. For example, an object with type int contains an integer value. The type is important because the collection of bits that represents one type of object will likely have a different value if interpreted as a different type of object. For example, the number 1 is represented in IEEE 754 (the IEEE Standard for Floating-Point Arithmetic) by the bit pattern 0x3f800000 (IEEE 754–2008). But if you were to interpret this same bit pattern as an integer, you’d get the value 1,065,353,216 instead of 1. Functions are not objects but do have types. A function type is characterized by both its return type as well as the number and types of its parameters. The C language also has pointers, which can be thought of as an address— a location in memory where an object or function is stored. A pointer type is derived from a function or object type called the referenced type. A pointer type derived from the referenced type T is called a pointer to T. Because objects and functions are different things, object pointers and function pointers are also different things, and should not be used interchangeably. In the following section, you’ll write a simple program that attempts to swap the values of two variables to help you better understand objects, functions, pointers, and types.

Declaring Variables When you declare a variable, you assign it a type and provide it a name, or identifier, by which to reference the variable. Listing 2-1 declares two integer objects with initial values. This simple program also declares, but doesn’t define, a swap function to swap those values. #include u void swap(int, int); // defined in Listing 2-2 int main(void) { int a = 21; int b = 17; v swap(a, b); printf("main: a = %d, b = %d\n", a, b); return 0; } Listing 2-1: Program meant to swap two integers

14   Chapter 2

This example program shows a main function with a single code block between the { } characters. This kind of code block is also known as a compound statement. We define two variables, a and b, within the main function. We declare the variables as having the type int and initialize them to 21 and 17, respectively. Each variable must have a declaration. The main function then calls the swap function v to try to swap the values of the two integers. The swap function is declared in this program u but not defined. We’ll look at some possible implementations of this function later in this section.

DECL A RING MULTIPLE VA RI A BLE S You can declare multiple variables in any single declaration, but doing so can get confusing if the variables are pointers or arrays, or the variables are different types. For example, the following declarations are all correct: char *src, c; int x, y[5]; int m[12], n[15][3], o[21]; The first line declares two variables, src and c, which have different types. The src variable has a type of char *, and c has a type of char. The second line again declares two variables, x and y, with different types. The variable x has a type int, and y is an array of five elements of type int. The third line declares three arrays— m, n, and o—with different dimensions and numbers of elements. These declarations are easier to understand if each is on its own line: char *src; // char c; // int x; // int y[5]; // int m[12]; // int n[15][3]; // int o[21]; //

src has c has a x has a y is an m is an n is an o is an

a type of char * type of char type int array of 5 elements of type int array of 12 elements of type int array of 15 arrays of 3 elements of type int array of 21 elements of type int

Readable and understandable code is less likely to have defects.

Swapping Values (First Attempt) Each object has a storage duration that determines its lifetime, which is the time during program execution for which the object exists, has storage, has a constant address, and retains its last-stored value. Objects must not be referenced outside their lifetime. Local variables such as a and b from Listing 2-1 have automatic storage duration, meaning that they exist until execution leaves the block in which they’re defined. We are going to attempt to swap the values stored in these two variables. Objects, Functions, and Types   15

Listing 2-2 is our first attempt to implement the swap function. void swap(int a, int b) { int t = a; a = b; b = t; printf("swap: a = %d, b = %d\n", a, b); } Listing 2-2: The swap function

The swap function declares two parameters, a and b, that you use to pass arguments to this function. C distinguishes between parameters, which are objects declared as part of the function declaration that acquire a value on entry to the function, and arguments, which are comma-separated expressions you include in the function call expression. We also declare a temporary variable t of type int in the swap function and initialize it to the value of a. This variable is used to temporarily save the value stored in a so that it is not lost during the swap. You can now compile and test the complete program by running the generated executable: % ./a.out swap: a = 17, b = 21 main: a = 21, b = 17

This result may be surprising. The variables a and b were initialized to 21 and 17, respectively. The first call to printf within the swap function shows these two values swapped, but the second call to printf in main shows the original values unchanged. Let’s examine what happened. C is a call-by-value (also called a pass-by-value) language, which means that when you provide an argument to a function, the value of that argument is copied into a distinct variable for use within the function. The swap function assigns the values of the objects you pass as arguments to the respective parameters. When the values of the parameters in the function are changed, the values in the caller are unaffected because they are distinct objects. Consequently, the variables a and b retain their original values in main during the second call to printf. The goal of the program was to swap the values of these two objects. By testing the program, we’ve discovered it has a bug, or defect.

Swapping Values (Second Attempt) To repair this bug, you can use pointers to rewrite the swap function. We use the indirection (*) operator to both declare pointers and dereference them, as shown in Listing 2-3. void swap(int *pa, int *pb) { int t = *pa; *pa = *pb;

16   Chapter 2

*pb = t; return; } Listing 2-3: The revised swap function using pointers

When used in a function declaration or definition, * acts as part of a pointer declarator indicating that the parameter is a pointer to an object or function of a specific type. In the rewritten swap function, we specify two parameters, pa and pb, and declare them both as type pointers to int. When you use the unary * operator in expressions within the function, the unary * operator dereferences the pointer to the object. For example, consider the following assignment: pa = pb;

This replaces the value of the pointer pa with the value of the pointer pb. Now consider the actual assignment in the swap function: *pa = *pb;

This dereferences the pointer pb, reads the referenced value, dereferences the pointer pa, and then overwrites the value at the location referenced by pa with the value referenced by pb. When you call the swap function in main, you must also place an ampersand (&) character before each variable name: swap(&a, &b);

The unary & is the address-of operator, which generates a pointer to its operand. This change is necessary because the swap function now accepts pointers to objects of type int as parameters instead of simply values of type int. Listing 2-4 shows the entire swap program with emphasis on the objects created during execution of this code and their values. #include void swap(int *pa, int *pb) { // pa → a: 21 pb → b: 17 int t = *pa; // t: 21 *pa = *pb; // pa → a: 17 pb → b: 17 *pb = t; // pa → a: 17 pb → b: 21 } int main(void) { int a = 21; // a: 21 int b = 17; // b: 17 swap(&a, &b); printf("a = %d, b = %d\n", a, b); // a: 17 b: 21 return 0; } Listing 2-4: Simulated call-by-reference Objects, Functions, and Types   17

Upon entering the main block, the variables a and b are initialized to 21 and 17, respectively. The code then takes the addresses of these objects and passes them to the swap function as arguments. Within the swap function, the parameters pa and pb now both declared to have the type pointer to int and contain copies of the arguments passed to swap from the calling function (in this case, main). These address copies still refer to the exact same objects, so when the values of the objects they reference are swapped in the swap function, the contents of the original objects declared in main are accessed and also swapped. This approach simulates callby-reference (also known as pass-by-reference) by generating object addresses, passing those by value, and then dereferencing the copied addresses to access the original objects.

Scope Objects, functions, macros, and other C language identifiers have scope that delimits the contiguous region where they can be accessed. C has four types of scope: file, block, function prototype, and function. The scope of an object or function identifier is determined by where it is declared. If the declaration is outside any block or parameter list, the identifier has file scope, meaning the scope is the entire text file in which it appears as well as any files included after that point. If the declaration appears inside a block or within the list of parameters, it has block scope, meaning that the identifier it declares is accessible only within the block. The identifiers for a and b from Listing 2-4 have block scope and can be used to refer to only these variables within the code block in the main function in which they’re defined. If the declaration appears within the list of parameter declarations in a function prototype (not part of a function definition), the identifier has function prototype scope, which terminates at the end of the function declarator. Function scope is the area between the opening { of a function definition and its closing }. A label name is the only kind of identifier that has function scope. Labels are identifiers followed by a colon and identify a statement in a function to which control may be transferred. Chapter 5 covers labels and control transfer. Scopes can be nested, with inner and outer scopes. For example, you can have a block scope inside another block scope, and every block scope is defined within a file scope. The inner scope has access to the outer scope, but not vice versa. As the name implies, any inner scope must be completely contained within the outer scopes that encompass it. If you declare the same identifier in both the inner scope and an outer scope, the identifier declared in the outer scope is hidden by the identifier within the inner scope, which takes precedence. In this case, naming the identifier will refer to the object in the inner scope; the object from the outer scope is hidden and cannot be referenced by its name. The easiest way to prevent this from becoming a problem is to use different names.

18   Chapter 2

Listing 2-5 demonstrates different scopes and how identifiers declared in inner scopes can hide identifiers declared in outer scopes. int j; // file scope of j begins void f(int i) { // block scope of i begins int j = 1; // block scope of j begins; hides file-scope j i++; // i refers to the function parameter for (int i = 0; i < 2; i++) { // block scope of loop-local i begins int j = 2; // block scope of the inner j begin; hides outer j printf("%d\n", j); // inner j is in scope, prints 2 } // block scope of the inner i and j ends printf("%d\n", j); // the outer j is in scope, prints 1 } // the block scope of i and j ends void g(int j); // j has function prototype scope; hides file-scope j Listing 2-5: Scope

There is nothing wrong with this code, provided the comments accurately describe your intent. Best practice is to use different names for different identifiers to avoid confusion, which leads to bugs. Using short names such as i and j is fine for identifiers with small scopes. Identifiers in large scopes should have longer, descriptive names that are unlikely to be hidden in nested scopes. Some compilers will warn about hidden identifiers.

Storage Duration Objects have a storage duration that determines their lifetime. Altogether, four storage durations are available: automatic, static, thread, and allocated. You’ve already seen that objects of automatic storage duration are declared within a block or as a function parameter. The lifetime of these objects begins when the block in which they’re declared begins execution, and ends when execution of the block ends. If the block is entered recursively, a new object is created each time, each with its own storage. N O T E Scope

and lifetime are entirely different concepts. Scope applies to identifiers, whereas lifetime applies to objects. The scope of an identifier is the code region where the object denoted by the identifier can be accessed by its name. The lifetime of an object is the time period for which the object exists. Objects declared in file scope have static storage duration. The lifetime of these objects is the entire execution of the program, and their stored value is initialized prior to program startup. You can also declare a variable within a block scope to have static storage duration by using the storageclass specifier static, as shown in the counting example in Listing 2-6. These objects persist after the function has exited. void increment(void) { static unsigned int counter = 0; Objects, Functions, and Types   19

counter++; printf("%d ", counter); } int main(void) { for (int i = 0; i < 5; i++) { increment(); } return 0; } Listing 2-6: A counting example

This program outputs 1 2 3 4 5. We initialize the static variable ­counter to 0 once at program startup, and increment it each time the i­ ncrement function is called. The lifetime of counter is the entire execution of the program, and it will retain its last-stored value throughout its lifetime. You could achieve the same behavior by declaring counter with file scope. However, it is good software engineering practice to limit the scope of an object wherever possible. Static objects must be initialized with a constant value and not a variable: int *func(int i) { const int j = i; // ok static int k = j; // error return &k; }

A constant value refers to literal constants (for example, 1, 'a', or 0xFF), enum members, and the results of operators such as alignof or sizeof; not const-qualified objects. Thread storage duration is used in concurrent programming and is not covered by this book. Allocated storage duration deals with dynamically allocated memory and is discussed in Chapter 6.

Alignment Object types have alignment requirements that place restrictions on the addresses at which objects of that type may be allocated. An alignment represents the number of bytes between successive addresses at which a given object can be allocated. CPUs may have different behavior when accessing aligned data (for example, the data address is a multiple of the data size) versus unaligned data. Some machine instructions can perform multibyte accesses on nonword boundaries, but there may be a performance penalty. A word is a natural, fixed-sized unit of adata handled by the instruction set or the hardware of the processor. Some platforms cannot access unaligned memory. Alignment requirements may depend on the CPU word size (typically, 16, 32, or 64 bits). 20   Chapter 2

Generally, C programmers need not concern themselves with alignment requirements, because the compiler chooses suitable alignments for its various types. Dynamically allocated memory from malloc is required to be sufficiently aligned for all standard types, including arrays and structures. However, on rare occasions, you might need to override the compiler’s default choices; for example, to align data on the boundaries of the memory cache lines that must start at power-of-two address boundaries, or to meet other system-specific requirements. Traditionally, these requirements were met by linker commands, or by overallocating memory with malloc followed by rounding the user address upward, or similar operations involving other nonstandard facilities. C11 introduced a simple, forward-compatible mechanism for specifying alignments. Alignments are represented as values of the type size_t. Every valid alignment value is a nonnegative integral power of two. An object type imposes a default alignment requirement on every object of that type: a stricter alignment (a larger power of two) can be requested using the alignment specifier (_Alignas). You can include an alignment specifier in the declaration specifiers of a declaration. Listing 2-7 uses the alignment specifier to ensure that good_buff is properly aligned (bad_buff may have incorrect alignment for member-access expressions). struct S { int i; double d; char c; }; int main(void) { unsigned char bad_buff[sizeof(struct S)]; _Alignas(struct S) unsigned char good_buff[sizeof(struct S)]; struct S *bad_s_ptr = (struct S *)bad_buff; // wrong pointer alignment struct S *good_s_ptr = (struct S *)good_buff; // correct pointer alignment } Listing 2-7: Use of the _Alignas keyword

Alignments are ordered from weaker to stronger (also called stricter) alignments. Stricter alignments have larger alignment values. An address that satisfies an alignment requirement also satisfies any valid, weaker alignment requirement.

Object Types This section introduces the object types in C. Specifically, we’ll cover the Boolean type, character types, and numerical types (including both integer and floating-point types).

Boolean Types Objects declared as _Bool can store only the values 0 and 1. This Boolean type was introduced in C99, and starts with an underscore to differentiate it in Objects, Functions, and Types   21

existing programs that had already declared their own identifiers named bool or boolean. Identifiers that begin with an underscore and either an uppercase letter or another underscore are always reserved. The idea is that the C Standards committee can create new keywords such as _Bool, assuming that you have avoided the use of reserved identifiers. If you haven’t, as far as the C Standards committee is concerned, it is your fault for not reading the standard carefully. If you include the header , you can also spell this type as bool and assign it the values true (which expands to the integer constant 1) and false (which expands to the integer constant 0). Here we declare two Boolean variables using both spellings of the type name: #include _Bool flag1 = 0; bool flag2 = false;

Both spellings will work, but it is better to use bool, as this is the longterm direction for the language.

Character Types The C language defines three character types: char, signed char, and unsigned char. Each compiler implementation will define char to have the same alignment, size, range, representation, and behavior as either signed char or unsigned char. Regardless of the choice made, char is a separate type from the other two and is incompatible with both. The char type is commonly used to represent character data in C language programs. In particular, objects of type char must be able to represent the minimum set of characters required in the execution environment (known as the basic execution character set), including upper- and lowercase letters, the 10 decimal digits, the space character, and various punctuation and control characters. The char type is inappropriate for integer data; it is safer to use signed char to represent small signed integer values, and unsigned char to represent small unsigned values. The basic execution character set suits the needs of many conventional data processing applications, but its lack of non-English letters is an obstacle to acceptance by international users. To address this need, the C Standards committee specified a new, wide type to allow large character sets. You can represent the characters of a large character set as wide characters by using the wchar_t type, which generally takes more space than a basic character. Typically, implementations choose 16 or 32 bits to represent a wide character. The C Standard Library provides functions that support both narrow and wide character types.

Numerical Types C provides several numerical types that can be used to represent integers, enumerators, and floating-point values. Chapter 3 covers some of these in more detail, but here’s a brief introduction. 22   Chapter 2

Integer Types Signed integer types can be used to represent negative numbers, positive numbers, and zero. The signed integer types include signed char, short int, int, long int, and long long int. Except for int itself, the keyword int may be omitted in the declarations for these types, so you might, for example, declare a type by using long long instead of long long int. For each signed integer type, there is a corresponding unsigned integer type that uses the same amount of storage: unsigned char, unsigned short int, unsigned int, unsigned long int, and unsigned long long int. The unsigned types can be used to represent only positive numbers and zero. The signed and unsigned integer types are used to represent integers of various sizes. Each platform (current or historical) determines the size for each of these types, given some constraints. Each type has a minimum representable range. The types are ordered by width, guaranteeing that wider types are at least as large as narrower types so that an object of type long long int can represent all values that an object of type long int can represent, an object of type long int can represent all values that can be represented by an object of type int, and so forth. The actual size of the various integer types can be inferred from the minimum and maximum representable values for the various integer types specified in the header file. The int type usually has the natural size suggested by the architecture of the execution environment, so the size would be 16 bits wide on a 16-bit architecture, and 32 bits wide on a 32-bit architecture. You can specify actualwidth integers by using type definitions from the or headers, like uint32_t. These headers also provide type definitions for the widest available integer types: uintmax_t and intmax_t. Chapter 3 covers integer types in excruciating detail. enum Types An enumeration, or enum, allows you to define a type that assigns names (enumerators) to integer values in cases with an enumerable set of constant values. The following are examples of enumerations: enum day { sun, mon, tue, wed, thu, fri, sat }; enum cardinal_points { north = 0, east = 90, south = 180, west = 270 }; enum months { jan = 1, feb, mar, apr, may, jun, jul, aug, sep, oct, nov, dec };

If you don’t specify a value to the first enumerator with the = operator, the value of its enumeration constant is 0, and each subsequent enumerator without an = adds 1 to the value of the previous enumeration constant. Consequently, the value of sun in the day enumeration is 0, mon is 1, and so forth. You can also assign specific values to each enumerator, as shown by the cardinal_points enumeration. Using = with enumerators may produce enumeration constants with duplicate values, which can be a problem if you

Objects, Functions, and Types   23

incorrectly assume all the values are unique. The months enumeration sets the first enumerator at 1, and each subsequent enumerator that isn’t specifically assigned a value will be incremented by 1. The actual value of the enumeration constant must be representable as an int, but its type is implementation defined. For example, Visual C++ uses a signed int, and GCC uses an unsigned int. Floating-Point Types The C language supports three floating-point types: float, double, and long double. Floating-point arithmetic is similar to, and often used as a model for, the arithmetic of real numbers. The C language supports a variety of floatingpoint representations including, on most systems, the IEEE Standard for Floating-Point Arithmetic (IEEE 754–2008). The choice of floating-point representation is implementation dependent. Chapter 3 covers floatingpoint types in detail.

void Types The void type is a rather strange type. The keyword void (by itself) means “cannot hold any value.” For example, you can use it to indicate that a function doesn’t return a value, or as the sole parameter of a function to indicate that the function takes no arguments. On the other hand, the derived type void * means that the pointer can reference any object. I’ll discuss derived types later in this chapter.

Function Types Function types are derived types. In this case, the type is derived from the return type and the number and types of its parameters. The return type of a function cannot be an array type. When you declare a function, you use the function declarator to specify the name of the function and the return type. If the declarator includes a parameter type list and a definition, the declaration of each parameter must include an identifier, except parameter lists with only a single parameter of type void, which needs no identifier. Here are a few function type declarations: int f(void); int *fip(); void g(int i, int j); void h(int, int);

First, we declare a function f with no parameters that returns an int. Next, we declare a function fip with no specified parameters that returns a pointer to an int. Finally, we declare two functions, g and h, each returning void and taking two parameters of type int. Specifying parameters with identifiers (as done here with g) can be problematic if an identifier is a macro. However, providing parameter 24   Chapter 2

names is good practice for self-documenting code, so omitting the identifiers (as done with h) is not typically recommended. In a function declaration, specifying parameters is optional. However, failing to do so is occasionally problematic. If you were to write the function declaration for fip in C++, it would declare a function accepting no arguments and returning an int *. In C, fip declares a function accepting any number of arguments of any type and returning an int *. You should never declare functions with an empty parameter list in C. First, this is a deprecated feature of the language that may be removed in the future. Second, the code could be ported to C++, so explicitly list parameter types and use void when there are no parameters. A function type with a parameter type list is known as a function prototype. A function prototype informs the compiler about the number and types of parameters a function accepts. Compilers use this information to verify that the correct number and type of parameters are used in the function definition and any calls to the function. The function definition provides the actual implementation of the function. Take a look at the following function definition: int max(int a, int b) { return a > b ? a : b; }

The return type specifier is int; the function declarator is max(int a, int b); and the function body is { return a > b ? a : b; }. The specification of a function type must not include any type qualifiers (see “Type Qualifiers” on page 32). The function body itself uses the condition operator (? :), which is explained further in Chapter 4. This expression states that if a is greater than b, return a; otherwise, return b.

Derived Types Derived types are types that are constructed from other types. These include pointers, arrays, type definitions, structures, and unions, all of which we’ll cover here.

Pointer Types A pointer type is derived from the function or object type that it points to, called the referenced type. A pointer provides a reference to an entity of the referenced type. The following three declarations declare a pointer to int, a pointer to char, and a pointer to void: int *ip; char *cp; void *vp;

Objects, Functions, and Types   25

Earlier in the chapter, I introduced the address-of (&) and indirection (*) operators. You use the & operator to take the address of an object or function. If the object is an int, for example, the result of the operator has the type pointer to int: int i = 17; int *ip = &i;

We declare the variable ip as a pointer to int and assign it the address of i. You can also use the & operator on the result of the * operator: ip = &*ip;

Dereferencing ip by using the indirection operator resolves to the actual object i. Taking the address of *ip by using the & operator retrieves the pointer, so these two operations cancel each other out. The unary * operator converts a pointer to a type into a value of that type. It denotes indirection and operates only on pointers. If the operand points to a function, the result of using the * operator is the function designator, and if it points to an object, the result is a value of the designated object. For example, if the operand is a pointer to int, the result of the indirection operator has the type int. If the pointer is not pointing to a valid object or function, bad things may happen.

Arrays An array is a contiguously allocated sequence of objects that all have the same element type. Array types are characterized by their element types and the number of elements in the array. Here we declare an array of 11 elements of type int identified by ia, and an array of 17 elements of type pointer to float identified by afp: int ia[11]; float *afp[17];

You use square brackets ([]) to identify an element of an array. For example, the following contrived code snippet creates the string "0123456789" to demonstrate how to assign values to the elements of an array: char str[11]; for (unsigned int i = 0; i < 10; ++i) { u str[i] = '0' + i; } str[10] = '\0';

The first line declares an array of char with a bound of 11. This allocates sufficient storage to create a string with 10 characters plus a null character.

26   Chapter 2

The for loop iterates 10 times, with the values of i ranging from 0 to 9. Each iteration assigns the result of the expression '0' + i to str[i]. Following the end of the loop, the null character is copied to the final element of the array str[10]. In the expression at u, str is automatically converted to a pointer to the first member of the array (an object of type char), and i has an unsigned integer type. The subscript ([]) operator and addition (+) operator are defined so that str[i] is identical to *(str + i). When str is an array object (as it is here), the expression str[i] designates the ith element of the array (counting from 0). Because arrays are indexed starting at 0, the array char str[11] is indexed from 0 to 10, with 10 being the last element, as referenced on the last line of this example. If the operand of the unary & operator is the result of a [] operator, the result is as if the & operator were removed and the [] operator were changed to a + operator. For example, &str[10] is the same as str + 10. You can also declare multidimensional arrays. Listing 2-8 declares arr in the function main as a two-dimensional 5 × 3 array of type int, also referred to as a matrix. void func(int arr[5]); int main(void) { unsigned int i = 0; unsigned int j = 0; int arr[3][5]; u func(arr[i]); v int x = arr[i][j]; return 0; } Listing 2-8: Matrix operations

More precisely, arr is an array of three elements, each of which is an array of five elements of type int. When you use the expression arr[i] at u (which is equivalent to *(arr+i)), the following occurs: 1. arr is converted to a pointer to the initial array of five elements of type int starting at arr[i]. 2. i is scaled to the type of arr by multiplying i by the size of one array of five int objects. 3. The results from steps 1 and 2 are added. 4. Indirection is applied to the result to produce an array of five elements of type int. When used in the expression arr[i][j] at v, that array is converted to a pointer to the first element of type int, so arr[i][j] produces an object of type int.

Objects, Functions, and Types   27

T Y PE DE FINITIONS You use a typedef to declare an alias for an existing type; it never creates a new type. For example, each of the following declarations creates a new type alias: typedef unsigned int uint_type; typedef signed char schar_type, *schar_p, (*fp)(void); On the first line, we declare uint_type as an alias for the type unsigned int. On the second line, we declare schar_type as an alias for signed char, schar_p as an alias for signed char *, and fp as an alias for signed char(*) (void). Identifiers that end in _t in the standard headers are type definitions (aliases for existing types). Generally speaking, you should not follow this convention in your own code because the C Standard reserve identifiers that match the patterns int[0-9a-z_]*_t and uint[0-9a-z_]*_t, and the Portable Operating System Interface (POSIX) reserves all identifiers that end in _t. If you define identifiers that use these names, they may collide with names used by the implementation, which can cause problems that are difficult to debug.

Structures A structure type (also known as a struct) contains sequentially allocated member objects. Each object has its own name and may have a distinct type— unlike arrays, which must all be of the same type. Structures are similar to record types found in other programming languages. Listing 2-9 declares an object identified by sigline that has a type of struct sigrecord and a pointer to the sigline object identified by sigline_p. struct sigrecord { int signum; char signame[20]; char sigdesc[100]; } sigline, *sigline_p; Listing 2-9: struct sigrecord

The structure has three member objects: signum is an object of type int, signame is an array of type char consisting of 20 elements, and sigdesc is an array of type char consisting of 100 elements. Structures are useful for declaring collections of related objects and may be used to represent things such as a date, customer, or personnel record. They are especially useful for grouping objects that are frequently passed together as arguments to a function, so you don’t need to repeatedly pass individual objects separately.

28   Chapter 2

Once you have defined a structure, you’ll likely want to reference its members. You reference members of an object of the structure type by using the structure member (.) operator. If you have a pointer to a structure, you can reference its members with the structure pointer (->) operator. Listing 2-10 demonstrates the use of each operator. sigline.signum = 5; strcpy(sigline.signame, "SIGINT"); strcpy(sigline.sigdesc, "Interrupt from keyboard"); u sigline_p = &sigline; sigline_p->signum = 5; strcpy(sigline_p->signame, "SIGINT"); strcpy(sigline_p->sigdesc, "Interrupt from keyboard"); Listing 2-10: Referencing structure members

The first three lines of Listing 2-10 directly access members of the sigline object by using the . operator. At u, we assign the pointer to sigline_p to the address of the sigline object. In the final three lines of the program, we indirectly access the members of the sigline object by using the -> operator through the sigline_p pointer.

Unions Union types are similar to structures, except that the memory used by the member objects overlaps. Unions can contain an object of one type at one time, and an object of a different type at a different time, but never both objects at the same time, and are primarily used to save memory. Listing 2-11 shows the union u that contains three structures: n, ni, and nf. This union might be used in a tree, graph, or other data structure that has some nodes that contain integer values (ni) and other nodes that contain floating-point values (nf). union { struct { int type; } n; struct { int type; int intnode; } ni; struct { int type; double doublenode; } nf; } u; u.nf.type = 1; u.nf.doublenode = 3.14; Listing 2-11: Unions Objects, Functions, and Types   29

As with structures, you can access union members via the . operator. Using a pointer to a union, you can reference its members with the -> operator. In Listing 2-11, the type member in the nf struct of the union is referenced as u.nf.type, and the doublenode member is referenced as u.nf  .doublenode. Code that uses this union will typically check the type of the node by examining the value stored in u.n.type and then accessing either the intnode or doublenode struct depending on the type. If this had been implemented as a structure, each node would contain storage for both the intnode and the doublenode members. The use of a union allows the same storage to be used for both members.

Tags Tags are a special naming mechanism for structs, unions, and enumerations. For example, the identifier s appearing in the following structure is a tag: struct s { //---snip--};

By itself, a tag is not a type name and cannot be used to declare a variable (Saks 2002). Instead, you must declare variables of this type as follows: struct s v; // instance of struct s struct s *p; // pointer to struct s

The names of unions and enumerations are also tags and not types, meaning that they cannot be used alone to declare a variable. For example: enum day { sun, mon, tue, wed, thu, fri, sat }; day today; // error enum day tomorrow; // OK

The tags of structures, unions, and enumerations are defined in a separate namespace from ordinary identifiers. This allows a C program to have both a tag and another identifier with the same spelling in the same scope: enum status { ok, fail }; // enumeration enum status status(void); // function

You can even declare an object s of type struct s: struct s s;

This may not be good practice, but it is valid in C. You can think of struct tags as type names and define an alias for the tag by using a typedef. Here’s an example: typedef struct s { int x; } t;

30   Chapter 2

This now allows you to declare variables of type t instead of struct s. The tag name in struct, union, and enum is optional, so you can just dispense with it entirely: typedef struct { int x; } t;

This works fine except in the case of self-referential structures that contain pointers to themselves: struct tnode { int count; struct tnode *left; struct tnode *right; };

If you omit the tag on the first line, the compiler may complain because the referenced structure on lines 3 and 4 has not yet been declared, or because the whole structure is not used anywhere. Consequently, you have no choice but to declare a tag for the structure, but you can declare a typedef as well: typedef struct tnode { int count; struct tnode *left; struct tnode *right; } tnode;

Most C programmers use a different name for the tag and the typedef, but the same name works just fine. You can also define this type before the structure so that you can use it to declare the left and right members that refer to other objects of type tnode: typedef struct tnode tnode; struct tnode { int count; tnode *left tnode *right; } tnode;

Type definitions can improve code readability beyond their use with structures. For example, all three of the following declarations of the signal function specify the same type: typedef void fv(int), (*pfv)(int); void (*signal(int, void (*)(int)))(int); fv *signal(int, fv *); pfv signal(int, pfv);

Objects, Functions, and Types   31

Type Qualifiers All the types examined so far have been unqualified types. Types can be qualified by using one or more of the following qualifiers: const, volatile, and restrict. Each of these qualifiers changes behaviors when accessing objects of the qualified type. The qualified and unqualified versions of types can be used interchangeably as arguments to functions, return values from functions, and members of unions. N O T E The _Atomic

type qualifier, available since C11, supports concurrent programs.

const Objects declared with the const qualifier (const-qualified types) are not modifiable. In particular, they’re not assignable but can have constant initializers. This means objects with const-qualified types can be placed in read-only memory by the compiler, and any attempt to write to them will result in a runtime error: const int i = 1; // const-qualified int i = 2; // error: i is const-qualified

It’s possible to accidentally convince your compiler to change a constqualified object for you. In the following example, we take the address of a const-qualified object i and tell the compiler that this is actually a pointer to an int: const int i = 1; // object of const-qualified type int *ip = (int *)&i; *ip = 2; // undefined behavior

C does not allow you to cast away the const if the original was declared as a const-qualified object. This code might appear to work, but it’s defective and may fail later. For example, the compiler might place the const-qualified object in read-only memory, causing a memory fault when trying to store a value in the object at runtime. C allows you to modify an object that is pointed to by a const-qualified pointer by casting the const away, provided that the original object was not declared const: int i const const const *(int *(int

32   Chapter 2

= 12; int j = 12; int *ip = &i; int *jp = &j; *)ip = 42; // ok *)jp = 42; // undefined behavior

volatile Objects of volatile-qualified types serve a special purpose. Static volatilequalified objects are used to model memory-mapped input/output (I/O) ports, and static constant volatile-qualified objects model memory-mapped input ports such as a real-time clock. The values stored in these objects may change without the knowledge of the compiler. For example, every time the value from a real-time clock is read, it may change, even if the value has not been written to by the C program. Using a volatile-qualified type lets the compiler know that the value may change, and ensures that every access to the real-time clock occurs (­otherwise, an access to the real-time clock may be optimized away or replaced by a previously read and cached value). In the following code, for example, the compiler must generate instructions to read the value from port and then write this value back to port: volatile int port; port = port;

Without the volatile qualification, the compiler would see this as a no-op (a programming statement that does nothing) and potentially eliminate both the read and the write. Also, volatile-qualified types are used for communications with signal handlers and with setjmp/longjmp (refer to the C Standard for information on signal handlers and setjmp/longjmp). Unlike in Java and other programming languages, volatile-qualified types in C should not be used for synchronization between threads.

restrict A restrict-qualified pointer is used to promote optimization. Objects ­indirectly accessed through a pointer frequently cannot be fully optimized because of potential aliasing, which occurs when more than one pointer refers to the same object. Aliasing can inhibit optimizations, because the compiler can’t tell if portions of an object can change values when another apparently unrelated object is modified, for example. The following function copies n bytes from the storage referenced by q to the storage referenced by p. The function parameters p and q are both restrict-qualified pointers: void f(unsigned int n, int * restrict p, int * restrict q) { while (n-- > 0) { *p++ = *q++; } }

Because both p and q are restrict-qualified pointers, the compiler can assume that an object accessed through one of the pointer parameters is not also accessed through the other. The compiler can make this Objects, Functions, and Types   33

assessment based solely on the parameter declarations without analyzing the function body. Although using restrict-qualified pointers can result in more efficient code, you must ensure that the pointers do not refer to overlapping memory to prevent undefined behavior.

Exercises Try these code exercises on your own: 1. Add a retrieve function to the counting example from Listing 2-6 to retrieve the current value of counter. 2. Declare an array of three pointers to functions and invoke the appropriate function based on an index value passed in as an argument.

Summary In this chapter, you learned about objects and functions and how they differ. You learned how to declare variables and functions, take the addresses of objects, and dereference those object pointers. You also learned about most of the object types that are available to C programmers as well as derived types. We’ll return to these types in later chapters to explore in more detail how they can be best used to implement your designs. In the next chapter, I provide detailed information about the two kinds of arithmetic types: integers and floating-point.

34   Chapter 2

3 ARITHME TIC T YPES

In this chapter, you’ll learn about the two kinds of arithmetic types: integers and floating-point types. Most operators in C operate on arithmetic types. Because C is a system-level language, performing arithmetic operations correctly can be difficult—resulting in frequent defects. This is partially because arithmetic operations in digital systems with limited range and precision do not always produce the same result as they would in ordinary mathematics. Being able to perform basic arithmetic correctly in C is an essential foundation to becoming a professional C programmer.

We’ll dive deep into how arithmetic works in the C language so that you have a firm grasp of these fundamental concepts. We’ll also look at how to convert one arithmetic type to another, which is necessary for performing operations on mixed types.

Integers As mentioned in Chapter 2, each integer type represents a finite range of integers. Signed integer types represent values that can be negative, zero, or positive; unsigned integers represent values that can be only zero or positive. The range that each type of integer can represent depends on your implementation. The value of an integer object is the ordinary mathematical value stored in the object. The representation of a value for an integer object is the particular encoding of the value in the bits of the object’s allocated storage. We’ll look at the representation in more detail later.

Padding and Precision All integer types except char, signed char, and unsigned char may contain unused bits, called padding, that allow implementations to accommodate hardware quirks (such as skipping over a sign bit in the middle of a multiple-word representation) or to optimally align with a target architecture. The number of bits used to represent a value of a given type, excluding ­padding but including the sign, is called the width and is often denoted by N. The precision is the number of bits used to represent values, excluding sign and padding bits.

The Header File The header file provides the minimum and maximum representable values for the various integer types. A representable value is one that can be represented in the number of bits available to an object of a particular type. Values that cannot be represented will be diagnosed by the compiler or converted to a representable but different (incorrect) value. Compiler writers provide the correct minimum, maximum, and width values for their implementations. To write portable code, you should use these constants, rather than integer literals such as +2147483647 that represent a specific limit and may change when porting to a different implementation. The C Standard imposes only three constraints on integer sizes. First, storage for every data type occupies an integral number of adjacent unsigned char objects (which may include padding). Second, each integer type has to support the minimum ranges, allowing you to depend on a portable range of values across any implementation. Third, smaller types cannot be wider than larger types. So, for example, USHRT_MAX cannot be greater than UINT_MAX, but they can be the same width.

36   Chapter 3

Declaring Integers Unless explicitly declared as unsigned, integer types are assumed to be signed (except for char, which the implementation can define as either a signed or unsigned integer type). The following are valid declarations of unsigned integers: unsigned unsigned unsigned unsigned

int ui; // unsigned is required u; // int can be omitted long long ull2; // int can be omitted char uc; // unsigned is required

When declaring signed integer types, you can omit the signed keyword—except for signed char, which requires the keyword to distinguish signed char from plain char. Unless it is the only keyword present, int can also be omitted. For example, instead of declaring a variable to be of type signed long long int, it is common practice to just declare it as long long and save some typing. The following are all valid declarations of signed integers: int i; // signed can be omitted long long int sll; // signed can be omitted long long sll2; // signed and int can be omitted signed char sc; // signed is required

Unsigned Integers Unsigned integers have ranges that start at 0, and their upper bound is greater than that of the corresponding signed integer type. Unsigned integers are frequently used for counting items that may have large, nonnegative quantities. Representation Unsigned integer types are easier to understand and to use than signed integer types. They represent values using a pure binary system with no offset: the least significant bit has the weight 20, the next least significant has the weight 21, and so forth. The value of the binary number is the sum of all the weights for the set bits. Table 3-1 shows some examples of unsigned values using an unpadded 8-bit representation. Table 3-1: 8-Bit Unsigned Values

Decimal

Binary

Hexadecimal

0

0000 0000

0x00

1

0000 0001

0x01

17

0001 0001

0x11

255

1111 1111

0xFF

Arithmetic Types   37

Unsigned integer types do not require the sign to be represented and so generally provide 1 bit greater precision than the corresponding signed integer types. Unsigned integer values range from 0 to a maximum value that depends on the width of the type. This maximum value is 2N – 1, where N is the width. For example, most x86 architectures use 32-bit integers with no padding bits, so an object of type unsigned int has a range of 0 to 232 – 1 (4,294,967,295). The constant expression UINT_MAX from specifies the implementation-defined upper range for this type. Table 3-2 shows the constant expressions from for each unsigned type, the minimum range required by the standard, and the actual range on modern x86 implementations. Table 3-2: Unsigned Integer Ranges

Constant expression

Minimum magnitudes

x86

Maximum value for an object of type

UCHAR_MAX

255 // 28 – 1

Same

unsigned char 

USHRT_MAX

16

65,535 // 2 – 1

Same

unsigned short int 

UINT_MAX

65,535 // 216 – 1

4,294,967,295

unsigned int

Same

unsigned long int

Same

unsigned long   long int

32

ULONG_MAX

4,294,967,295 // 2

–1

ULLONG_MAX

18,446,744,073,709,551,615 // 264 – 1  

Wraparound Wraparound occurs when you perform arithmetic operations that result in values too small (less than 0) or too large (greater than 2N – 1) to be represented as a particular unsigned integer type. In this case, the value is reduced modulo the number that is one greater than the largest value that can be represented in the resulting type. Wraparound is well-defined behavior in the C language. Whether it is a defect in your code depends on the context. If you are counting something and the value wraps, it is likely to be an error. However, the use of wraparound in certain encryption algorithms is intentional. For example, the code in Listing 3-1 initializes ui to its maximum value and then increments it. The resulting value cannot be represented as an unsigned int, so it wraps around to 0. If this value is then decremented, it falls outside the range once more, so it wraps around again to UINT_MAX. unsigned int ui = UINT_MAX; // 4,294,967,295 on x86 ui++; printf("ui = %u\n", ui); // ui is 0 ui--; printf("ui = %u\n", ui); // ui is 4,294,967,295 Listing 3-1: Unsigned integer wraparound

38   Chapter 3

Because of wraparound, an unsigned integer expression can never evaluate to less than 0. It’s easy to lose track of this and implement comparisons that are always true or always false. For example, the i in the following for loop can never take on a negative value, so this loop will never terminate: for (unsigned int i = n; i >= 0; --i)

This behavior has caused some notable real-world bugs. For example, all six power-generating systems on a Boeing 787 are managed by a corresponding generator control unit. Boeing’s laboratory testing discovered that an internal software counter in the generator control unit wraps around after running continuously for 248 days, according to the Federal Aviation Administration.6 This defect causes all six generator control units on the engine-mounted generators to enter fail-safe mode at the same time. To avoid unplanned behavior (such as having your airplane fall from the sky), it’s important to check for wraparound by using the limits from . You should be careful when implementing these checks, because it is easy to make mistakes. For example, the following code contains a defect as sum + ui can never be larger than UINT_MAX: extern unsigned int ui, sum; // assign values to ui and sum if (sum + ui > UINT_MAX) too_big(); else sum = sum + ui;

If the result of sum + ui is larger than UINT_MAX, it’s reduced modulo UINT_MAX + 1. Therefore, this entire test is useless, and the generated code will unconditionally perform the summation. Quality compilers might issue a warning pointing this out, but not all do. To remedy this, we can subtract sum from both sides of the inequality to form the following effective test: extern unsigned int ui, sum; // assign values to ui and sum if (ui > UINT_MAX - sum) too_big(); else sum = sum + ui; UINT_MAX is the largest value that can be represented as an unsigned int, and sum is a value between 0 and UINT_MAX. If sum is equal to UINT_MAX, the result of the subtraction is 0, and if sum is equal to 0, the result of the subtraction is UINT_MAX. Because the result of this operation will always fall in the allowable range of 0 to UINT_MAX, it can never wrap.

6. S ee Airworthiness Directives; The Boeing Company Airplanes, https://www.federalregister.gov/d/2015-10066/. Arithmetic Types   39

The same problem occurs when checking the result of an arithmetic operation against 0, the minimum unsigned value: extern unsigned int i, j; // assign values to i and j if (i - j < 0) // cannot happen negative(); else i = i - j;

Because unsigned integer values can never be negative, the subtraction will be performed unconditionally. Quality compilers may warn about this mistake as well. Instead of this useless test, we can check for wraparound by testing whether j is greater than i: if (j > i)  // correct negative(); else i = i - j;

If j > i, the result would wrap around, so the possibility of wraparound is clearly detected. By eliminating the subtraction operation in the test, we eliminate the possibility of wraparound occurring during the test. WARNING

Keep in mind that the width used when wrapping depends on the implementation, which means you can obtain different results on different platforms. Unless you take this into account, your code won’t be portable.

Signed Integers Each unsigned integer type (excluding _Bool) has a corresponding signed integer type that occupies the same amount of storage. We use signed integers to represent negative, zero, and positive values, the range of which depends on the number of bits allocated to the type and the representation. Representation Representing signed integer types is more complicated than representing unsigned integer types. Historically, the C language has supported three signed integer representation schemes: Sign and magnitude  The high-order bit indicates the sign, and the remaining bits represent the magnitude of the value in pure binary notation. Ones’ complement  The sign bit is given the weight −(2N − 1 − 1), and the other value bits have the same weights as for unsigned. Two’s complement  The sign bit is given the weight −(2N − 1), and the other value bits have the same weights as for unsigned. 

40   Chapter 3

You cannot choose which representation to use; that is determined by the implementers of C for the various systems. Though all three are still in use, two’s complement is by far the most common representation—so much so that the C Standards committee intends, starting with C2x, to accommodate only two’s-complement representation. The remainder of this book assumes two’s-complement representation. Signed integer types with a width of N can represent any integer value in the range of –2N – 1 to 2N–1– 1. This means, for example, that an 8-bit value of type signed char has a range of –128 to 127. Two’s complement can represent an additional most negative value. The most negative value for an 8-bit signed char is –128, and its absolute value |–128| cannot be represented as this type. This leads to some interesting edge cases, which we’ll examine later in this chapter and the next. Table 3-3 shows the constant expressions from for each signed type, the minimum range required by the standard, and the actual range on modern x86 implementations. Table 3-3: Signed Integer Ranges

Constant expression Minimum magnitudes 7

–127 // – (2 – 1)

SCHAR_MIN

7

+127 // 2 – 1

SCHAR_MAX

15

–32,767 // –(2 – 1)

SHRT_MIN

15

x86

Type

–128

signed char

Same

signed char

–32,768

short int

SHRT_MAX

+32,767 // 2 – 1

Same

short int

INT_MIN

–32,767 // –(215 – 1)

–2,147,483,648

int

+2,147,483,647

int

15

+32,767 // 2 – 1

INT_MAX

31

LONG_MIN

–2,147,483,647 // –(2

–2,147,483,648

long int

LONG_MAX

+2,147,483,647 // 231 – 1

– 1)

Same

long int

LLONG_MIN

–9,223,372,036,854,775,807 // – (263 – 1)

-9,223,372,036,854,775,808

long long int

LLONG_MAX

+9,223,372,036,854,775,807 // 263 – 1

Same

long long int

The two’s complement representation for negative signed values includes a sign bit and other value bits. The sign bit is given the weight –(2N – 1). To negate a value in two’s-complement representation, simply toggle each nonpadding bit and then add 1 (with carries as necessary), as shown in Figure 3-1. 0

0

1

0

1

0

0

1

1

1

0

1

0

1

1

0

+

1

=

0

0

1

0

1

0

0

1

1

1

0

1

0

1

1

1

Figure 3-1: Negating an 8-bit value in two’s-complement representation

Arithmetic Types   41

Table 3-4 shows the binary and decimal representations for an 8-bit two’s-complement signed integer type with no padding (that is, N = 8). This is not required knowledge, but as a C programmer, you will likely find it useful. Table 3-4: 8-Bit Two’s-Complement Values

Binary

Decimal

Weighting

0

0

00000001

1

20

01111110

126

26 + 25 + 24 + 23 + 22 + 21

01111111

127

2N − 1 − 1

00000000

10000000 10000001 11111110 11111111

−128 −127 −2 −1

N−1

)+0

N−1

)+1

−(2 −(2

N−

−(2

SCHAR_MAX SCHAR_MIN

1 ) + 126

N−1

−(2

Constant

) + 127

Overflow Overflow occurs when a signed integer operation results in a value that cannot be represented in the resulting type. For example, the following implementation of a function-like macro that returns the absolute value of an integer can overflow: // undefined or wrong for most negative value #define Abs(i) ((i) < 0 ? -(i) : (i))

We’ll examine macros in detail in Chapter 9. For now, you can think of function-like macros as functions that operate on generic types. On the surface, this macro appears to correctly implement the absolute value function by returning the nonnegative value of i without regard to its sign. We use the conditional (?:) operator (which I’ll cover in more detail in the next chapter) to test whether the value of i is negative. If so, i is negated to -(i); otherwise, it evaluates to the unmodified value (i). Because we’ve implemented Abs as a function-like macro, it can take an argument of any type. Of course, invoking this macro with an unsigned integer is pointless, because unsigned integers can never be negative, so the macro’s output would just reproduce the argument. However, we can invoke the function with a variety of signed integer and floating-point types, as in the following invocation: signed int si = -25; signed int abs_si = Abs(si); printf("%d\n", abs_si); // prints 25

42   Chapter 3

In this example, we pass an object of type signed int with the value –25 as an argument to the Abs macro. This invocation expands to the following: signed int si = -25; signed int abs_si = ((si) < 0 ? –(si) : (si)); printf("%d\n", abs_si); // prints 25

The macro correctly returned the absolute value of –25. So far, so good. The problem is that the negative of the two’s-complement most negative value for a given type cannot be represented in that type, so this use of the Abs function results in signed integer overflow. Consequently, this implementation of Abs is defective and can do anything, including unexpectedly returning a negative value: signed int si = INT_MIN; signed int abs_si = Abs(si); // undefined behavior printf("%d\n", abs_si);

So, what should Abs(INT_MIN) return to fix this behavior? Signed integer overflow is undefined behavior in C, allowing implementations to silently wrap (the most common behavior), trap, or both. Traps interrupt execution of the program so that no further operations are performed. Common architectures like x86 do a combination of both. Because the behavior is undefined, no universally correct solution to this problem exists, but we can at least test for the possibility of undefined behavior before it occurs and take appropriate action. To make the absolute-value macro useful for a variety of types, we’ll add a type-dependent flag argument to it. The flag represents the *_MIN macro, which matches the type of the first argument. This value is returned in the problem case: #define AbsM(i, flag) ((i) >= 0 ? (i) : ((i)==(flag) ? (flag) : -(i))) signed int si = -25; // try INT_MIN to trigger the problem case signed int abs_si = AbsM(si, INT_MIN); if (abs_si == INT_MIN) goto recover; // special case else printf("%d\n", abs_si); // prints 25

The AbsM macro tests for the most negative value and simply returns it if found instead of triggering the undefined behavior by negating it. On some systems, the C Standard library implements the following intonly absolute-value function to avoid overflow when the function is passed INT_MIN as an argument: int abs(int i) { return (i >= 0) ? i : -(unsigned)i; // avoids overflow }

Arithmetic Types   43

In this case, i is converted to an unsigned int and negated. I’ll discuss conversions in more detail later in this chapter. Perhaps surprisingly, unary minus (-) operator is defined for unsigned integer types. The resulting unsigned integer value is reduced modulo the number that is one greater than the largest value that can be represented by the resulting type. Finally, i is implicitly converted back to signed int as required by the return statement. Because -INT_MIN can’t be represented as a signed int, the result is implementation defined. This is why this implementation is used only on some systems, and even on these systems, the abs function returns an incorrect value. The Abs and AbsM implementations use function-like macros to evaluate their parameters more than once. This can cause surprises when the arguments cause program state to change. These are called side effects and are covered in detail in the next chapter. Function calls, on the other hand, evaluate each argument only once. Unsigned integers have well-defined wraparound behavior. Signed integer overflow, or the possibility of it, should always be considered a defect.

Integer Constants Integer constants (or integer literals) are constants we use to introduce particular integer values into a program. For example, you might use them in a declaration or assignment to initialize a counter to 0. C has three kinds of integer constants that use different number systems: decimal constants, octal constants, and hexadecimal constants. Decimal constants always begin with a nonzero digit. For example, the following code uses two decimal constants: unsigned int ui = 71; int si; si = -12;

In this example code, we initialize ui to the decimal constant 71 and assign si the decimal constant value -12. Use decimal constants when introducing regular integer values into your code. If a constant starts with a 0, optionally followed by digits 0 through 7, it is an octal constant. Here’s an example: int agent = 007; int permissions = 0777;

In this example, 007 octal equals 7 decimal, and the octal constant 0777 equals the decimal value 511. Octal constants are convenient when dealing with 3-bit fields, for example. You can also create a hexadecimal constant by prefixing a sequence of decimal digits and the letters a (or A) through f (or F) with 0x or 0X. For example: int burger = 0xDEADBEEF;

44   Chapter 3

Use hexadecimal constants when the constant you are introducing is meant to represent a bit pattern more than a particular value; for example, when representing an address. Idiomatically, most hexadecimal constants are written like 0xDEADBEEF because it resembles a typical hex dump. It’s probably a good idea for you to write all your hexadecimal constants like this. You can also append a suffix to your constant to specify its type. Without a suffix, a decimal constant is given the int type if it can be represented as a value in that type. If it cannot be represented as an int, it will be represented as a long int or long long int. The suffixes are U for unsigned, L for signed long, and LL for long long. These can be combined. For example the ULL suffix represents the unsigned long long type. Here are some examples: unsigned int ui = 71U; signed long int sli = 9223372036854775807L; unsigned long long int ui = 18446744073709551615ULL;

If we don’t use a suffix, and the integer constant isn’t of the required type, it may be implicitly converted. (We’ll discuss implicit conversion in “Arithmetic Conversion” on page 49.) This may result in a surprising conversion or a compiler diagnostic, so it’s best to correctly specify the desired type of your integer constants. Section 6.4.4.1 of the C Standard contains more information on integer constants.

Floating-Point Floating-point is the most common representation for real numbers in c­ omputers. Floating-point representation is a technique that uses ­scientific notation to encode numbers with a base number and an exponent. For example, the decimal number 123.456 can be represented as 1.23456 × 102 while the binary number 0b10100.110 can be represented as 1.0100110 × 24. You can generate floating-point representations in several ways. The C Standard doesn’t require that implementations use any specific model, although it does require every implementation to support some model. To keep things simple, we’ll assume conformance to Annex F (the most common floating-point format). You can test the values of the __STDC_IEC_559__ or __STDC_IEC_60559_BFP__ macros in newer compilers to determine whether the implementation conforms to Annex F. This section explains floating-point types, arithmetic, values, and constants, so you will know how and when to use them to emulate math on real numbers and when to avoid them.

Arithmetic Types   45

Floating-Point Types C has three floating-point types: float, double, and long double. The float type can be used for floating-point calculations in which the result can be adequately represented as a single-precision result. The common IEC 60559 float type encodes values using 1 sign bit, 8 exponent bits, and 23 significand bits (ISO/IEC/IEEE 60559:2011). The double type provides greater precision but requires additional storage. It encodes values using 1 sign bit, 11 exponent bits, and 52 significand bits. These types are illustrated in Figure 3-2. s

exp

significand

1

8

23 32 bits

s

exp

significand

1

11

52 64 bits

Figure 3-2: The float and double types

Each implementation assigns the long double type one of the following formats: •

IEC 60559 quadruple (or binary128) format7

• • •

IEC 60559 binary64-extended format A non-IEC 60559 extended format IEC 60559 double (or binary64) format

Recommended practice for compiler implementers is to match the long double type with the IEC 60559 binary128 format or an IEC 60559 binary64extended format. IEC 60559 binary64-extended formats include the common 80-bit IEC 60559 format. Larger types have greater precision but require more storage. Any value that can be represented as a float can also be represented as a double, and any value that can be represented as a double can be represented as a long double. Future versions of the standard may include additional floating-point types, including those with greater range, precision, or both than long double, or lesser range and precision, such as 16-bit floating-point type.

7. IEC 60559 added binary128 to its basic formats in the 2011 revision.

46   Chapter 3

SIGN, E X PONE NT, A ND SIGNIFIC A ND As with integers, the sign bit represents whether the number is positive or negative: 0 denotes a positive number, and 1 denotes a negative number. The exponent field needs to represent both positive and negative exponents. To avoid storing the exponent as a signed number, a bias is implicitly added to the actual exponent to get the stored exponent. For the float type, the bias is 127. Consequently, to express an exponent of 0, we would store 127 in the exponent field. A stored value of 200 indicates an exponent of 200 – 127, or 73. Exponents of –127 (where every exponent bit is 0) and +128 (where every exponent bit is 1) are reserved for special numbers. Similarly, the bias for double-precision numbers is 1023. This means that the value stored will range from 0 to 255 for a float, and 0 to 2047 for a double. The significand bits represent the precision bits of the number. For example, if you were to represent the value 1.0100110 × 24 as a floating-point value, the significand refers to the precision bits 1.0100110, and the exponent refers to the power of 2, which is 4 in this example (Hollasch 2019).

Floating-Point Arithmetic Floating-point arithmetic is similar to, and used to model, the arithmetic of real numbers. However, there are differences to consider. In particular, unlike the arithmetic of real numbers, floating-point numbers are bounded in magnitude and have finite precision. Addition and multiplication operations are not associative, the distributive property doesn’t hold, nor do many other properties that are valid for real numbers. Floating-point types cannot represent all real numbers exactly, even when they can be represented in a small number of decimal digits. For example, common decimal constants such as 0.1 can’t be represented exactly as binary floating-point numbers. Floating-point types may lack the necessary precision for various applications such as loop counters or performing financial calculations. See CERT C rule FLP30-C (Do not use floating-point variables as loop counters) for more information.

Floating-Point Values Ordinarily, all of the significand bits in a floating-point type express significant figures, in addition to a leading 1, which is implied and omitted, though still considered part of the value. As a special case, to represent the value 0, the exponent and significand must both be 0; zeros are signed (+0 and –0) according to the sign bit, so there are two floating-point zero values: a positive one and a negative one. There are no leading zeros in the significand of a normal floating-point value; leading zeros are removed by adjusting the exponent. Therefore, float has 24 significant bits of precision, double has 53 significant bits of precision, and long double has 113 significant bits of precision (assuming the Arithmetic Types   47

quadruple 128-bit IEC 60559 format). These are normalized numbers, and they preserve the full precision of the significand. Nonnormalized (or subnormal) numbers are very small positive and negative numbers (but not 0) whose representation would result in an exponent that is less than the smallest representable value. Figure 3-3 is a number line showing the range of subnormal values around 0. A nonzero number represented with the minimum exponent (that is, the implicit 1 bit is taken to be 0) is a subnormal, even if all its explicit significand bits are 1. The precision of denormalized floating-point values is less than that of normalized floating-point values.

-FLT_MIN

0

FLT_MIN

Figure 3-3: Domain of subnormal numbers

Floating-point types can also represent values that are not floatingpoint numbers, such as negative and positive infinity and not-a-number (NaN) values. NaNs are values that do not represent a real number. Having infinity available as a specific value allows operations to continue past overflow situations, which often produces the desired result without requiring special treatment. For example, dividing any positive or negative nonzero value by either positive or negative zero8 yields either positive or negative infinity. Operations with infinite values are well-defined in the IEEE floating-point standard. A quiet NaN propagates through almost every arithmetic operation without raising a floating-point exception and is typically tested after a selected sequence of operations. A signaling NaN generally raises a floating-point exception immediately when occurring as an arithmetic operand. Floatingpoint exceptions are an advanced topic not covered here. For more information, refer to Annex F of the C Standard. The NAN and INFINITY macros and the nan functions in provide designations for IEC 60559 quiet NaNs and infinities. The SNANF, SNAN, and SNANL macros (ISO/IEC TS 18661-1:2014, ISO/IEC TS 18661-3:2015) defined in provide designations for IEC 60559 signaling NaNs. The C Standard does not require full support for signaling NaNs. One way to identify the kind of floating-point value you’re dealing with is to use the fpclassify function-like macro, which classifies its argument value as NaN, infinite, normal, subnormal, or zero. #include int fpclassify(real-floating x);

In Listing 3-2, we use the fpclassify macro in the show_classification function to determine whether a floating-point value of type double is a normal value, subnormal value, zero, infinity, or NaN. 8. Distinct values −0 and +0 compare as equal.

48   Chapter 3

const char *show_classification(double x) { switch(fpclassify(x)) { case FP_INFINITE: return "Inf"; case FP_NAN: return "NaN"; case FP_NORMAL: return "normal"; case FP_SUBNORMAL: return "subnormal"; case FP_ZERO: return "zero"; default: return "unknown"; } Listing 3-2: The fpclassify macro

The function argument x (a double in this example) is passed to the fpclassify macro, which switches on the result. The function returns a string corresponding to the class of value stored in x.

Floating-Point Constants A floating-point constant is a decimal or hexadecimal number that represents a signed real number. You should use floating-point constants to represent floating-point values that cannot be changed. The following are some examples of floating-point constants: 15.75 1.575E1 /* 1575e-2 /* -2.5e-3 /* 25E-4 /*

15.75 */ 15.75 */ -0.0025 */ 0.0025 */

All floating-point constants have a type. The type is double if unsuffixed, float if suffixed by the letter f or F, or long double if suffixed by the letter l or L, as shown here: 10.0 /* type double */ 10.0F /* type float */ 10.0L /* type long double */

Arithmetic Conversion Frequently, a value represented in one type (for example, float) must be represented in a different type (for example, int). This might occur, for example, when you have an object of type float and need to pass it as an argument to a function that accepts only an object of type int. When such conversions are necessary, you should always ensure that the value is representable in the new type. I’ll discuss this further in “Safe Conversions” on page 54. Values can be implicitly or explicitly converted from one arithmetic type to another. You can use the cast operator to perform explicit conversions. Listing 3-3 shows two examples of casts.

Arithmetic Types   49

int si = 5; short ss = 8; long sl = (long)si;u unsigned short us = (unsigned short)(ss + sl);v Listing 3-3: Cast operators

To perform a cast, place a type name in parentheses just before the expression. The cast converts the expression to the unqualified version of the type name in parentheses. Here, we cast the value of si to the type long u. Because si is of type int, this cast (from a smaller to a larger integer type of the same signedness) is guaranteed to be safe because the value can always be represented in the larger type. The second cast in this code snippet v casts the result of the expression (ss + sl) to type unsigned short. Because the value is converted to an unsigned type (unsigned short) with less precision, the result of the conversion might not be equal to the original value. (Some compilers might warn about this; others won’t.) In this example, the result of the expression (13) can be correctly represented in the resulting type. Implicit conversion, also known as coercion, occurs automatically in expressions as required. This happens, for example, when operations are performed on mixed types. In Listing 3-3, implicit conversions are used to convert ss to the type of sl so that the addition ss + sl can be performed on a common type. The rules concerning which values are implicitly converted to which types are somewhat complicated and involve three concepts: integer conversion rank, integer promotions, and the usual arithmetic conversions. We’ll discuss these in the following sections.

Integer Conversion Rank An integer conversion rank is a standard rank ordering of integer types used to determine a common type for computations. Every integer type has an integer conversion rank that determines when and how conversions are implicitly performed. The C Standard, section 6.3.1.1, paragraph 1 (ISO/IEC 9899:2018), states that every integer type has an integer conversion rank where the following applies: • • •

• • 50   Chapter 3

No two signed integer types have the same rank, even if they have the same representation. The rank of a signed integer type is greater than the rank of any signed integer type with less precision. The rank of long long int is greater than the rank of long int, which is greater than the rank of int, which is greater than the rank of short int, which is greater than the rank of signed char. The rank of any unsigned integer type equals the rank of the corresponding signed integer type, if any. The rank of char equals the rank of signed char and unsigned char.

• •



The rank of _Bool is less than the rank of all other standard integer types. The rank of any enumerated type equals the rank of the compatible integer type. Each enumerated type is compatible with char, a signed integer type, or an unsigned integer type. The rank of any extended signed integer type relative to another extended signed integer type with the same precision is implementation-defined but still subject to the other rules for determining the integer conversion rank.

Integer Promotions A small type is an integer with a lower conversion rank than int or unsigned int. Integer promotion is the process of converting values of small types to an int or unsigned int. Integer promotions allow you to use an expression of a small type in any expression where an int or unsigned int may be used. For example, you could use the lower-ranked integer type—typically, char or short—on the right-hand side of an assignment or as an argument to a function. Integer promotions serve two primary purposes. First, they encourage operations to be performed in a natural size (int) for the architecture, which improves performance. Second, they help avoid arithmetic errors from the overflow of intermediate values, as shown in the following code snippet: signed char cresult, c1, c2, c3; c1 = 100; c2 = 3; c3 = 4; cresult = c1 * c2 / c3;

Without integer promotion, c1 * c2 would result in an overflow of the signed char type on platforms where signed char is represented by an 8-bit two’s-complement value, because 300 is outside the range of values that can be represented in an object of this type, which is –128 to 127. However, because of integer promotion, c1, c2, and c3 are implicitly converted to objects of type signed int, and the multiplication and division operations take place in this size. There is no possibility of overflow while performing these operations, because the resulting values can always be represented (objects of type signed int have a range of −2N – 1 to 2N − 1 – 1). In this specific example, the result of the entire expression is 75, which is within range of the signed char type, so the value is preserved when stored in cresult. Prior to the first C Standard, compilers used one of two approaches to integer promotions: the unsigned preserving approach or the value preserving approach. In the unsigned preserving approach, the compiler promotes unsigned small types to unsigned int. In the value preserving approach, if all values of the original type can be represented as an int, the value of the original small type will be converted to int. Otherwise, it is converted to unsigned int. When developing the original version of the standard (C89), the C Standards committee decided on value preserving rules, because they Arithmetic Types   51

produce incorrect results less often than the unsigned preserving approach. If necessary, you can override this behavior by using explicit type casts, as we did in Listing 3-3. The result of promoting small unsigned types depends on the precision of the integer types, which is implementation-defined. For example, the x86 architecture has an 8-bit char type, a 16-bit short type, and a 32-bit int type. For implementations that target this architecture, values of both unsigned char and unsigned short are promoted to signed int because all the values that can be represented in these smaller types can be represented as a signed int. However, 16-bit architectures, such as Intel 8086/8088 and the IBM Series/1, have an 8-bit char type, a 16-bit short type, and a 16-bit int type. For implementations that target these architectures, values of type unsigned char are promoted to signed int, while values of type unsigned short are promoted to unsigned int. This is because all the values that can be represented as an 8-bit unsigned char type can be represented as a 16-bit signed int, but some values that can be represented as a 16-bit unsigned short cannot be represented as a 16-bit signed int.

Usual Arithmetic Conversions The usual arithmetic conversions are rules for yielding a common type by balancing both operands of a binary operator to a common type, or balancing the second and third arguments of the conditional (? :) operator to a common type. Balancing conversions changes one or both operands of different types to the same type. Many operators that accept integer operands—including *, /, %, +, -, , =, ==, != , &, ^, |, and ? :—perform conversions using the usual arithmetic conversions. The usual arithmetic conversions are applied to the promoted operands. The usual arithmetic conversions first check whether one of the operands in the balancing conversion is a floating-point type. If so, it applies the following rules: 1. If one type of either operand is long double, the other operand is converted to long double. 2. Otherwise, if one type of either operand is double, the other operand is converted to double. 3. Otherwise, if the type of either operand is float, the other operand is converted to float. 4. Otherwise, the integer promotions are performed on both operands. If one operand has the type double and the other operand has the type int, for example, the operand of type int is converted to an object of type double. If one operand has the type float and the other operand has the type double, the operand of type float is converted to an object of type  double. If neither operand is a floating-point type, the following usual arithmetic conversion rules are applied to the promoted integer operands: 52   Chapter 3

1. If both operands have the same type, no further conversion is needed. 2. Otherwise, if both operands have signed integer types or both have unsigned integer types, the operand with the type that has the lesser integer conversion rank is converted to the type of the operand with greater rank. If one operand has the type int and the other operand has the type long, for example, the operand of type int is converted to an object of type long. 3. Otherwise, if the operand that has the unsigned integer type has a rank greater than or equal to the rank of the other operand’s type, then the operand with the signed integer type is converted to the type of the operand with the unsigned integer type. For example, if one operand has the type signed int, and the other operand has the type unsigned int, the operand of type signed int is converted to an object of type unsigned int. 4. Otherwise, if the type of the operand with the signed integer type can represent all of the values of the type of the operand with unsigned integer type, then the operand with unsigned integer type is converted to the type of the operand with signed integer type. For example, if one operand has the type unsigned int and the other operand has the type signed long long, and the signed long long type can represent all the values of the unsigned int type, then the operand of type unsigned int is converted to an object of type signed long long. This is the case for implementations with a 32-bit int type and a 64-bit long long type, such as x86-32 and x86-64. 5. Otherwise, both operands are converted to the unsigned integer type corresponding to the type of the operand with signed integer type. These conversion rules, which evolved as new types were added in C’s early years, take some getting used to. The irregularities in these patterns resulted from varying architectural properties (notably, the PDP-11’s automatic promotion of char to int) coupled with a desire to avoid changing the behavior of existing programs, and (subject to those constraints) a desire for uniformity. When in doubt, use type casts to explicitly force the conversions that you intend. That said, try not to overuse explicit conversions because casts can disable important diagnostics.

An Example of Implicit Conversion The following example illustrates the use of integer conversion rank, integer promotions, and the usual arithmetic conversions. This code compares the signed char value c for equality with the unsigned int value ui. We’ll assume this code is being compiled for the x86 architecture: unsigned int ui = UINT_MAX; signed char c = -1; if (c == ui) { puts("-1 equals 4,294,967,295"); } Arithmetic Types   53

The variable c is of type signed char. Because signed char has a lower integer conversion rank than int or unsigned int, the value stored in c is promoted to an object of type signed int when used in the comparison. This is accomplished by sign-extending the original value of 0xFF to 0xFFFFFFFF. Sign extension is used to convert a signed value to a larger-width object. The sign bit is copied into each bit position of the expanded object. This operation preserves the sign and magnitude when converting a value from a smaller to a larger, signed integer type. Next, the usual arithmetic conversions are applied. Because the operands to the equal (==) operator have different signedness and equal rank, the operand with the signed integer type is converted to the type of the operand with the unsigned integer type. The comparison is then performed as a 32-bit unsigned operation. Because UINT_MAX has the same values as the promoted and converted value of c, the comparison yields 1, and the code snippet prints the following: -1 equals 4,294,967,295

Safe Conversions Both implicit and explicit conversions (the result of a cast operation) can produce values that can’t be represented in the resulting type. It’s preferable to perform operations on objects of the same type to avoid conversions. However, conversions are unavoidable when a function returns or accepts an object of a different type. In these cases, it is necessary to ensure that the conversion is performed correctly. Integer Conversions Integer conversions occur when a value of an integer type is converted to a different integer type. Conversions to larger types of the same signedness are always safe and don’t need to be checked. Most other conversions can produce unexpected results if the resulting value cannot be represented in the resulting type. To perform these conversions correctly, you must test that the value stored in the original integer type is within the range of values that can be represented in the resulting integer type. As an example, the do_stuff function shown in Listing 3-4 accepts a signed long argument value that needs to be used in a context in which only a signed char is appropriate. To perform this conversion safely, the function checks that value can be represented as a signed char in the range [SCHAR_MIN, SCHAR_MAX] and returns an error if it is not. #include #include errno_t do_stuff(signed long value) { if ((value < SCHAR_MIN) || (value > SCHAR_MAX)) {

54   Chapter 3

return ERANGE; } signed char sc = (signed char)value; // Cast quiets warning //---snip--} Listing 3-4: Safe conversion

The specific range tests vary based on the conversion. See CERT C rule INT31-C (Ensure that integer conversions do not result in lost or misinterpreted data) for more information. Integer to Floating-Point Conversions If the value of the integer type being converted to a floating-point type can be represented exactly in the new type, it is unchanged. If the value being converted is in the range of values that can be represented but not exactly, the result is rounded to either the nearest higher or nearest lower representable value, depending on the implementation. If the value being converted is outside the range of values that can be represented, the behavior is undefined. CERT C rule FLP36-C (Preserve precision when converting integral values to floating-point type) provides more information and examples of these conversions. Floating-Point to Integer Conversions When a finite value of a floating type is converted to an integer type (other than bool), the fractional part is discarded. If the value of the integral part cannot be represented by the integer type, the behavior is undefined. Floating-Point Demotions Converting a floating-point value to a larger floating-point type is always safe. Demoting a floating-point value (that is, converting to a smaller floating-point type) is similar to converting an integer value to a floating-point type. Floating-point types that conform to Annex F support signed infinity. Demoting values of floating-point types for these implementations will always succeed because all values are within range. See CERT C rule FLP34-C (Ensure that floating-point conversions are within range of the new type) for more information on floating-point conversions.

Summary In this chapter, you learned about integers and floating-point types. You also learned about implicit and explicit conversions, integer conversion rank, integer promotions, and the usual arithmetic conversions. In the next chapter, you’ll learn about operators and how to write simple expressions to perform operations on various object types.

Arithmetic Types   55

4 E X PR E SSIONS A ND OPE R ATORS

In this chapter, you’ll learn about operators and how to write simple expressions to perform operations on various object types. An operator is a keyword or one or more punctuation characters used to perform an operation. When an operator is applied to one or more operands, it becomes an expression that computes a value and that might have side effects. Expressions are sequences of operators and operands that compute a value or accomplish another purpose. The operands can be identifiers, constants, string literals, and other expressions.

In this chapter, we discuss simple assignment before stepping back to describe the mechanics of expressions (operators and operands, value computations, side effects, precedence, and order of evaluation). We then discuss specific operators including arithmetic, bitwise, cast, conditional, alignment, relational, compound assignment, and the comma operator. We’ve introduced many of these operators and expressions in previous chapters; here, we detail their behavior and how best to use them. Finally, the chapter ends with a discussion of pointer arithmetic.

Simple Assignment A simple assignment replaces the value stored in the object designated by the left operand with the right operand. The value of the right operand is converted to the type of the assignment expression. Simple assignment has three components: the left operand, the assignment (=) operator, and the right operand, as shown in the following example: int i = 21; // declaration with initializer int j = 7; // declaration with initializer i = j; // simple assignment

The first two lines are declarations that define and initialize i with the value 21, and j with the value 7. An initializer uses an expression but is not itself an assignment expression, because an initializer is always part of a declaration. The third line contains a simple assignment. You must define or declare all identifiers that appear in an expression such as simple assignment for your code to compile. In simple assignment, the rvalue is converted to the type of the lvalue and then stored in the object designated by the lvalue. In the assignment i = j, the value is read from j and written to i. Because both i and j are the same type (int), no conversion is necessary. The assignment expression has the value of the result of the assignment and the type of the lvalue. The left operand in simple assignment is always an expression (with an object type other than void), and we refer to it as an lvalue. The l in lvalue originally comes from it being the left operand, but it may be more correct to think of it as standing for locator value, because it must designate an object. In this example, the identifiers for both objects i and j are lvalues. An lvalue can also be an expression such as *(p+4), provided it references an object in memory. The right operand is also an expression but can simply be a value and does not need to identify an object. We refer to this value as an rvalue (right operand) or expression value. The rvalue does not need to refer to an object, as you can see in the following statement, which uses the types and values from the preceding example: j = i + 12; // j now has the value 19

58   Chapter 4

The expression i + 12 is not an lvalue, because there is no underlying object storing the result. Instead, i by itself is an lvalue that is automatically converted into an rvalue to be used as an operand to the addition operation. The resulting value from the addition operation (which has no memory location associated with it) is also an rvalue. C constrains where lvalues and rvalues may appear. The following statements illustrate the correct and incorrect use of lvalues and rvalues: int i = int 7 =

i; 5; // i is an lvalue, 5 is an rvalue j = i; // lvalues can appear on the right-hand side of an assignment i; // error: rvalues can't appear on the left-hand side of an assignment

The assignment 7 = i won’t work, because the rvalue must always go on the right side of the operator. In the following example, the right operand has a different type than the assignment expression, so the value of i is first converted to a signed char type. The value of the expression enclosed in parentheses is then converted to the long int type: signed char c; int i = INT_MAX; long k; k = (c = i);

Assignment must deal with real-world constraints. Specifically, simple assignment can result in truncation if a value is converted to a narrower type. As mentioned in Chapter 3, each object requires a fixed number of bytes of storage. The value of i can always be represented by k (a larger type of the same signedness). However, in this example, the value of i is converted to signed char (the type of the assignment expression c = i). The value of the expression enclosed in parentheses is then converted to the type of the outer assignment expression—that is, long int type. Assuming c has fewer bits available to represent the value stored in i, values greater than SCHAR_MAX are truncated, so the eventual value stored in k (−1) is truncated. To prevent values from being truncated, make sure that you choose a sufficiently wide type that can represent any value that might arise, or check for overflow.

Evaluations Now that we’ve looked at simple assignment, let’s step back for a moment and look at how expressions are actually evaluated. Evaluation mostly means simplifying an expression down to a single value. However, the evaluation of an expression can include both value computations and the initiation of side effects. A value computation is the calculation of the value that results from the evaluation of the expression. Computing the final value may involve determining the identity of the object or reading the value previously assigned Expressions and Operators   59

to an object. For example, the following expression contains several value computations to determine the identity of i, a, and a[i]: a[i] + f() + 9

Because f is a function and not an object, the expression f() does not involve determining the identity of f. The value computations of operands must occur before the value computation of the result of the operator. In this example, separate value computations read the value of a[i] and determine the value returned by the call to function f. A third computation then sums these values to obtain the value returned by the overall expression. If a[i] is an array of int, and f() returns an int, the result of the expression will have the type int. Side effects are changes to the state of the execution environment. Side effects include writing to an object, accessing (reading or writing) a volatile-qualified object, I/O, assignment, or calling a function that does any of these things. The previous example can be slightly modified to add an assignment. Updating the stored value of j is a side effect of the assignment: int j; j = a[i] + f() + 9;

The assignment to j is a side effect that changes the state of the execution environment. Depending on the definition of the f function, the call to f may also have side effects.

Function Invocation A function designator is an expression that has function type and is used to invoke a function. In the following function invocation, max is the function designator: int x = 11; int y = 21; int max_of_x_and_y = max(x, y);

The max function returns the larger of its two arguments. In an expression, a function designator is converted to pointer-to-function returning type at compile time. The value of each argument must be of a type that can be assigned to an object with (the unqualified version of) the type of its corresponding parameter. The number and type of the arguments need to agree with the number and type of the parameters accepted by the function. Here, that means two integer arguments. C also supports variadic functions, which can accept a variable number of arguments (the printf function is an example of a variadic function).

60   Chapter 4

We can also pass one function to another, as shown by Listing 4-1. int f(void) { // ---snip-- return 0; } void g(int (*func)(void)) { // ---snip-- if (func() != 0) printf("g failed\n"; // ---snip--} // ---snip--g(f); // call g with function-pointer argument // ---snip--Listing 4-1: Passing one function to another function

This code passes the address of a function designated by f to another function, g. The function g accepts a function pointer to a function that accepts no arguments and returns int. A function passed as an argument is implicitly converted to a function pointer. The definition of g makes this explicit; an equivalent declaration is void g(int func(void)).

Increment and Decrement Operators The increment (++) and decrement (--) operators increment and decrement a modifiable lvalue, respectively. Both are unary operators, because they take a single operand. These operators can be used as either prefix operators, which come before the operand, or postfix operators, which come after the operand. The prefix and postfix operators have different behaviors, which means they are commonly used as trick questions in quizzes and interviews. A prefix increment performs the increment before returning the value, whereas a postfix increment returns the value and then performs the increment. Listing 4-2 illustrates these behaviors by performing a prefix or postfix increment or decrement operation and then assigning the result to e. int int e = e = e = e =

i = 5; e; // i++; // i--; // ++i; // --i; //

result of the expression postfix increment: i has the value 6; e has the value 5 postfix decrement: i has the value 5; e has the value 6 prefix increment: i has the value 6; e has the value 6 prefix decrement: i has the value 5; e has the value 5

Listing 4-2: Prefix and postfix increment and decrement operators

The i++ operation in this example returns the unchanged value 5, which is then assigned to e. The value of i is then incremented as a side effect of the operation.

Expressions and Operators   61

The prefix increment operator increments the value of the operand, and the expression returns the new value of the operand after it has been incremented. Consequently, the expression ++i is equivalent to i = i + 1, except that i is evaluated only once. The ++i operation in this example returns the incremented value 6, which is then assigned to e.

Operator Precedence and Associativity In mathematics and computer programming, the order of operations (or operator precedence) is a collection of rules that dictate the order in which operations are performed in a given expression. For example, multiplication is normally granted a higher precedence than addition. Therefore, the expression 2 + 3 × 4 is interpreted to have the value 2 + (3 × 4) = 14, not (2 + 3) × 4 = 20. Associativity determines how operators of the same precedence are grouped when no explicit parentheses are used. If adjacent operators have equal precedence, the choice of which operation to apply first is determined by the associativity. Left-associative operators cause the operations to be grouped from the left while right-associative operators cause the operations to be grouped from the right. Grouping can be thought of as the implicit introduction of parentheses. For example, the addition (+) operator has left associativity, so the expression a + b + c is interpreted as ((a + b) + c). The assignment operator is right-associative, so the expression a = b = c is interpreted as (a = (b = c)). Table 4-1 lists the precedence and associativity of C operators, as specified by the language syntax.7 Operators are listed top to bottom, in descending precedence. Table 4-1: Operator Precedence and Associativity

Precedence

Operator

Description

Associativity

0

(...)

Forced grouping

Left

1

++ --

Postfix increment and decrement

Left

()

Function call

[]

Array subscripting

.

Structure and union member access

->

Structure and union member access through pointer

(type){list}

Compound literal

7. T  his table is derived from the C Operator Precedence table at the C++ Reference website, https://en.cppreference.com/w/c/language/operator_precedence.

62   Chapter 4

Precedence

Operator

Description

Associativity

2

++ --

Prefix increment and decrement

Right

+ -

Unary plus and minus

! ~

Logical NOT and bitwise NOT

(type)

Type cast

*

Indirection (dereference)

&

Address-of

sizeof

Size of

_Alignof

Alignment requirement

3

* / %

Multiplication, division, and remainder

4

+ -

Addition and subtraction

5

Bitwise left shift and right shift

6

=

Relational operators > and ≥

7

== !=

Equal to and not equal to

8

&

Bitwise AND

9

^

Bitwise XOR (exclusive or)

10

|

Bitwise OR (inclusive or)

11

&&

Logical AND

12

||

Logical OR

13

?:

Conditional operator

14

=

Simple assignment

+= -=

Assignment by sum and difference

*= /= %=

Assignment by product, quotient, and remainder

=

Assignment by bitwise left shift and right shift

&= ^= |=

Assignment by bitwise AND, XOR, and OR

,

Expression sequencing

15

Left

Right

Left

Sometimes operator precedence can be intuitive, and sometimes it can be misleading. For example, the postfix ++ and -- operators have higher precedence than both the prefix ++ and -- operators, which in turn have the same precedence as the unary * operator. Moreover, if p is a pointer, then *p++ is equivalent to *(p++), and ++*p is equivalent to ++(*p), because both the prefix ++ operator and the unary * operator are right-associative. If two operators have the same precedence and association, they are evaluated from left to right. Listing 4-3 illustrates the precedence rules among these operators. char abc[] = "abc"; char xyz[] = "xyz"; char *p = abc; Expressions and Operators   63

printf("%c", ++*p); p = xyz; printf("%c", *p++); Listing 4-3: Operator precedence

The pointer in the expression ++*p is first dereferenced, producing the character 'a'. This value is then incremented, resulting in the character 'b'. On the other hand, the pointer in the expression *p++ is incremented first, so it refers to the 'y' character. However, the result of postfix increment operators is the value of the operand so that the original pointer value is dereferenced, producing the 'x' character. Consequently, this code prints out the characters bx. You can use parentheses, ( ), to change or clarify the order of operations.

Order of Evaluation The order of evaluation of the operands of any C operator, including the order of evaluation of any subexpressions, is generally unspecified. The compiler will evaluate them in any order, and may choose a different order when the same expression is evaluated again. This latitude allows the compiler to produce faster code by choosing the most efficient order. The order of evaluation is constrained by operator precedence and associativity. Listing 4-4 demonstrates the order of evaluation for function arguments. We invoke the max function we defined earlier with two arguments, which are the result of calling functions f and g, respectively. The order of evaluation of the expressions passed to max is unspecified, meaning that f and g could be called in either order. int glob; // static storage initialized to 0 int f(void) { return glob + 10; } int g(void) { glob = 42; return glob; } int main(void) { int max_value = max(f(), g()); // ---snip--} Listing 4-4: Order of evaluation for function arguments

The global variable glob is accessed by both functions f and g, meaning they rely on shared state. When calculating their return value, the values passed as arguments to max may differ between compilations. If f is called first, it will return 10, but if it is called last, it will return 52. Function g always returns 42 regardless of the order of evaluation. Consequently, the 64   Chapter 4

max function (which returns the greater of the two values) may return either

42 or 52, depending on the order of evaluation of its arguments. The only sequencing guarantees provided by this code are that both f and g are called before max, and that the executions of f and g do not interleave. This code can be rewritten to ensure it always behaves in a predictable, portable manner: int f_val = f(); int g_val = g(); int max_value = max(f_val, g_val);

In this revised program, f is called to initialize the variable f_val. This is guaranteed to be sequenced before the execution of g, which is called in the subsequent declaration to initialize the variable g_val. If one evaluation is sequenced before another evaluation, the first evaluation must complete before the second evaluation can begin. You can use sequence points to guarantee, for example, that an object will be written before it is read as part of a separate evaluation. The execution of f is guaranteed to be sequenced before the execution of g because a sequence point exists between the evaluation of one full expression and the next full expression. Sequence points are discussed in detail in the following subsections.

Unsequenced and Indeterminately Sequenced Evaluations The executions of unsequenced evaluations can interleave, meaning that the instructions can be executed in any order, provided that the execution is sequentially consistent—that reads and writes are performed in the order specified by the program (Lamport 1979). Some evaluations are indeterminately sequenced, which means they cannot interleave but can still be executed in any order. For example, the following statement contains several value computations and side effects: printf("%d\n", ++i + ++j * --k);

The values of i, j, and k must be read before their values can be incremented or decremented. This means that the reading of i must be sequenced before the increment side effect, for example. Similarly, all side effects for the operands of the multiplication operation need to complete before the multiplication can occur. Finally, the multiplication has to complete before the addition because of operator precedence rules, and all side effects must complete for the operands of the addition operation before it can occur. These constraints produce a partial ordering among these operations, because they don’t require that j is incremented before k is decremented, for example. Unsequenced evaluations in this expression can be performed in any order. This allows the compiler to reorder operations and cache values in registers, allowing for faster overall execution. Function executions, on the other hand, are indeterminately sequenced and do not interleave with each other.

Expressions and Operators   65

Sequence Points A sequence point is the juncture at which all side effects will have completed. These are implicitly defined by the language, but you can control where they occur by the way you specify your program logic. The sequence points are enumerated in Annex C of the C Standard. A sequence point occurs between the evaluation of one full expression (an expression that is not part of another expression or declarator) and the next full expression to be evaluated. A sequence point also occurs upon entering or exiting a called function. If a side effect is unsequenced relative to either a different side effect on the same scalar or a value computation that uses the value of the same scalar object, the code has undefined behavior. A scalar type is either an arithmetic type or pointer type. The expression i++ * i++ in the following code snippet performs two unsequenced operations on i: int i = 5; printf("Result = %d\n", i++ * i++);

You might think this code will produce the value 30, but because this code has undefined behavior, this outcome isn’t guaranteed. Conservatively, we can ensure that side effects have completed before the value is read by placing every side-effecting operation in its own full expression. We can rewrite this code as follows to eliminate the undefined behavior: int i = 5; int j = i++; int k = i++; printf("Result = %d\n", j * k);

This code now contains a sequence point between every side-effecting operation. However, it’s impossible to tell whether this rewritten code represents the original intent of the programmer, because the original code had no defined meaning. If you choose to omit sequence points, you must be sure you completely understand the sequencing of side effects. This same code can also be written as follows without changing the behavior: int i = 5; int j = i++; printf("Result = %d\n", j * i++);

Now that we have described the mechanics of expressions, we will return to discussing specific operators, starting with the sizeof operator.

sizeof Operator You can use the sizeof operator to find the size in bytes of its operand; specifically, it returns an unsigned integer of size_t type that represents the size. Knowing the correct size of an operand is necessary for most memory 66   Chapter 4

operations, including allocating and copying storage. The size_t type is defined in as well as other header files. You’ll need to include one of these header files to compile any code that references size_t. You can pass the sizeof operator an unevaluated expression of a complete object type or a parenthesized name of such a type: int i; size_t i_size = sizeof i; // the size of the object i size_t int_size = sizeof(int); // the size of the type int

It is always safe to parenthesize the operand to sizeof, because parenthesizing an expression does not change the way the size of the operand is calculated. The result of invoking the sizeof operator is a constant expression unless the operand is a variable-length array. The operand to sizeof is not evaluated. If you need to determine the number of bits of storage available, you can multiply the size of an object by CHAR_BIT, which gives the number of bits contained in a byte. For example, the expression CHAR_BIT * sizeof(int) will produce the number of bits in an object of type int. Object types other than character types can include padding as well as value representation bits. Different target platforms can pack bytes into multiple-byte words in different ways, called endianness.8 All this variation implies that, for interhost communication, you should adopt a standard for the external format and use format conversion functions to marshal arrays of external data to and from multiple-byte native objects.

Arithmetic Operators Several operators that perform arithmetic operations on arithmetic types are detailed in the following sections. Some of these operators can also be used with nonarithmetic operands.

Unary + and – Operators The unary + and – operators operate on a single operand of arithmetic type. The – operator returns the negative of its operand (that is, it behaves as though the operand were multiplied by –1). The unary + operator just returns the value. These operators exist primarily to express positive and negative numbers. If the operand has a small integer type, it is promoted (see Chapter 3), and the result of the operation has the result of the promoted type. As a point of trivia, C has no negative integer constants; a value such as –25 is actually an rvalue of type int with the value 25 preceded by the unary – operator.

8. T his term is drawn from Jonathan Swift’s 1726 satire, Gulliver’s Travels, in which civil war erupts over whether the big end or the little end of a boiled egg is the proper end to crack open. Expressions and Operators   67

Logical Negation Operator The result of the unary logical negation (!) operator is as follows: • •

0 if the value of its operand is not 0 1 if the value of its operand is 0

The operand is a scalar type. The result has type int for historical reasons. The expression !E is equivalent to (0 == E). The logical negation operator is frequently used to check for null pointers; for example, !p is equivalent to (NULL == p).

Multiplicative Operators The binary multiplicative operators include multiplication (*), division (/), and remainder (%). The usual arithmetic conversions are implicitly performed on multiplicative operands to find a common type. You can multiply and divide both floating-point and integer operands, but remainder operates on only integer operands. Various programming languages implement different kinds of integer division operations, including Euclidean, flooring, and truncating. In Euclidean division, the remainder is always nonnegative (Boute 1992). In flooring division, the quotient is rounded toward negative infinity (Knuth 1997). In truncating division, the result of the / operator is the algebraic quotient with any fractional part discarded. This is often referred to as truncation toward zero. The C programming language implements truncating division, meaning that the remainder always has the same sign as the dividend, as shown in Table 4-2. Table 4-2: Truncating Division /

Quotient

%

Remainder

10 / 3

3

10 % 3

1

10 / –3

–3

10 % –3

1

–10 / 3

–3

–10 % 3

–1

–10 / –3

3

–10 % –3

–1

To generalize, if the quotient a / b is representable, then the expression (a / b) * b + a % b equals a. Otherwise, if the value of the divisor is equal to 0 or a / b overflows, both a / b and a % b will result in undefined behavior. It’s worth taking the time to understand the behavior of the % operator to avoid surprises. For example, the following code defines a faulty function called is_odd that attempts to test whether an integer is odd: bool is_odd(int n) { return n % 2 == 1; }

68   Chapter 4

Because the result of the remainder operation always has the sign of the dividend n, when n is negative and odd, n % 2 returns −1, and the function returns false. A correct, alternative solution is to test that the remainder is not 0 (because a remainder of 0 is the same regardless of the sign of the dividend): bool is_odd(int n) { return n % 2 != 0; }

Many CPUs implement remainder as part of the division operator, which can overflow if the dividend is equal to the minimum negative value for the signed integer type and the divisor is equal to −1. This occurs even though the mathematical result of such a remainder operation is 0. The C Standard Library provides floating-point remainder, truncation, and rounding functions, including fmod.

Additive Operators The binary additive operators include addition (+) and subtraction (−). Addi­ tion and subtraction can be applied to two operands of arithmetic types, but you can also use them to perform scaled pointer arithmetic. I’ll discuss pointer arithmetic near the end of this chapter; the discussion here is limited to operations on arithmetic types. The binary + operator sums its two operands. The binary - operator subtracts the right operand from the left operand. The usual arithmetic conversions are performed on operands of arithmetic type for both operations.

Bitwise Operators We use bitwise operators to manipulate the bits of an object or any integer expression. Typically, they’re used on objects that represent bitmaps: each bit indicates that something is “on” or “off,” “enabled” or “disabled,” or another binary pairing. Bitwise (| & ^ ~) operators treat the bits as a pure binary model without concern for the values represented by these bits: 1 1 0 1 = 13 ^ 0 1 1 0 = 6 = 1 0 1 1 = 11

Bitmaps are best represented as unsigned integer types, as the sign bit can be better used as a value within the bitmap, and operations on the ­values are less prone to undefined behavior.

Expressions and Operators   69

Complement Operator The unary complement (~) operator works on a single operand of integer type and returns the bitwise complement of its operand; that is, a value in which each bit of the original value is flipped. The complement operator is used in applying the POSIX umask, for example. A file’s permission mode is the result of a logical AND operation between the complement of the mask and the process’s requested permission mode setting. Integer promotions are performed on the operand, and the result has the promoted type. For example, the following code snippet applies the ~ operator to a value of type unsigned char: unsigned char uc = UCHAR_MAX; // 0xFF int i = ~uc;

On an architecture with an 8-bit char type and 32-bit int type, uc is assigned the value 0xFF. When uc is used as the operand to the ~ operator, uc is promoted to signed int by zero-extending it to 32 bits, 0x000000FF. The complement of this value is 0xFFFFFF00. Therefore, on this platform, complementing an unsigned short type always results in a negative value of type signed int. As a general policy and to avoid surprises such as this, all bitwise manipulations should involve a single, sufficiently wide unsigned integer type.

Shift Operators Shift operations shift the value of each bit of an operand of integer type by a specified number of positions. Shifting is commonly performed in system programming, where bit masks are common. They may also be used in code that manages network protocols or file formats to pack or unpack data. They include left-shift operations, of the form: shift-expression > additive expression

The shift expression is the value to be shifted, and the additive expression is the number of bits by which to shift the value. Figure 4-1 illustrates a logical left shift of 1 bit. 0

0

0

1

0

1

1

1

0

0

1

0

1

1

1

0

0

Figure 4-1: Logical left shift of 1 bit

The additive expression determines the number of bits by which to shift the value. For example, the result of E1 > E2 is the value of E1 right-shifted E2 bit positions. If E1 has an unsigned type or if E1 has a signed type and a nonnegative value, the value of the result is the integral part of the quotient of E1/2E2. If E1 has a signed type and a negative value, the resulting value is implementation defined and may be either an arithmetic (signextended) shift or a logical (unsigned) shift, as shown in Figure 4-2. 1

0

0

1

0

1

1

1

1

1

0

0

1

0

1

1

0

1

0

0

1

0

1

1

1

0

1

0

0

1

0

1

1

Figure 4-2: Arithmetic (signed) right shift and logical (unsigned) right shift of 1 bit

In both kinds of shifts, the integer promotions are performed on the operands, each of which has the integer type. The type of the result is that of the promoted left operand. The usual arithmetic conversions are not performed. Even though you could use left and right shifts to multiply or divide by a power of two, it’s not a good idea to use a shift operation for this purpose. It is best to use the multiplication and division operations and let the compiler determine when to optimize processing by replacing these operations with a shift operation. Performing this substitution yourself is an example of premature optimization. Donald Knuth, author of The Art of Computer Programming (1997), has described premature optimization as “the root of all evil.” The number of bits shifted should not be negative, greater than, or equal to the width of the promoted left operand, because doing so is undefined behavior. Listing 4-5 shows how to perform right-shift operations on signed and unsigned integers that are free from these errors. extern int si1, si2, sresult; extern unsigned int ui1, ui2, uresult; // ---snip--u if ( (si2 < 0) || (si2 >= sizeof(int)*CHAR_BIT) ) { /* error */ } else { sresult = si1 >> si2; } v if (ui2 >= sizeof(unsigned int)*CHAR_BIT) { /* error */ } else { uresult = ui1 >> ui2; } Listing 4-5: Correct right-shift operations Expressions and Operators   71

For signed integers, u you must ensure that the number of bits shifted is not negative, or greater than, or equal to the width of the promoted left operand. For unsigned integers, v you omit the test for negative values, as unsigned integers can never be negative. You can perform left-shift operations in a similar manner.

Bitwise AND Operator The binary bitwise AND (&) operator returns the bitwise AND of two operands of integer type. The usual arithmetic conversions are performed on both operands. Each bit in the result is set if and only if each of the corresponding bits in the converted operands is set, as shown in Table 4-3. Table 4-3: Bitwise AND Truth Table

x

y

x&y

0

0

0

0

1

0

1

0

0

1

1

1

Bitwise Exclusive OR Operator The bitwise exclusive OR (^) operator returns the bitwise exclusive OR of the operands of integer type. In other words, each bit in the result is set if and only if exactly one of the corresponding bits in the converted operands is set, as shown in Table 4-4. This is sometimes thought of as “one or the other, but not both.” Table 4-4: Bitwise Exclusive OR Truth Table

x

y

x^y

0

0

0

0

1

1

1

0

1

1

1

0

Exclusive OR is equivalent to the addition operation on the integers modulo 2—that is, because of wraparound 1 + 1 mod 2 = 0 (Lewin 2012). The operands must be integers, and the usual arithmetic conversions are performed on both. Beginners commonly mistake the exclusive OR operator for an exponent operator, erroneously believing that the expression 2 ^ 7 will compute 2 raised to the power of 7. The correct way to raise a number to a certain power in C is to use the pow functions defined in , as shown in

72   Chapter 4

Listing 4-6. The pow functions operate on floating arguments and return a floating result, so be aware that these functions might fail to produce the expected results because of truncation or other errors. #include #include int main(void) { int i = 128; if (i == pow(2, 7)) { puts("equal"); } } Listing 4-6: Using the pow functions

This code calls the pow function to compute 2 raised to the power of 7. Because 27 equals 128, and assuming 128 is exactly representable in the type double, this program will print equal.

Bitwise Inclusive OR Operator The bitwise inclusive OR (|) operator operator returns the bitwise inclusive OR of the operands of integer type. The operands must be integers, and the usual arithmetic conversions are performed on both. Each bit in the result is set if and only if at least one of the corresponding bits in the converted operands is set, as shown in Table 4-5. Table 4-5: Bitwise Inclusive OR Truth Table

x

y

x|y

0

0

0

0

1

1

1

0

1

1

1

1

Logical Operators The logical AND (&&) operator and logical (||) OR operator are used primarily for logically joining two or more expressions of scalar type. They’re commonly used in condition tests, such as in the first operand of the conditional operator, the controlling expression of an if statement, or the controlling expression of a for loop, to combine multiple comparisons together. You shouldn’t use logical operators with bitmap operands, as they are intended primarily for Boolean logic. The && operator returns 1 if neither of its operands is equal to 0, and returns 0 otherwise. Logically, this means that a && b is true only if both a is true and b is true. Expressions and Operators   73

The || operator returns 1 if either of its operands is not equal to 0, and returns 0 otherwise. Logically, this means that a || b is true if a is true, b is true, or both a and b are true. The C Standard defines both operations in terms of “not equal to zero” because the operands can have values other than 0 and 1. Both o ­ perators accept operands of scalar type (integers, floats, and pointers), and the result of the operation has type int. Unlike the corresponding bitwise binary operators, the logical AND operator and logical OR operator guarantee left-to-right evaluation; if the second operand is evaluated, there is a sequence point between the evaluations of the first and second operands. The logical operators short-circuit: the second operand is not evaluated if the result can be deduced solely by evaluating the first operand. For example, the expression 0 && unevaluated returns 0 regardless of the value of unevaluated because there is no possible value for unevaluated that produces a different result. Because of this, unevaluated is not evaluated to determine its value. The same is true for 1 || unevaluated because this expression will always return 1. Short-circuiting is commonly used in operations with pointers: bool isN(int* ptr, int n){ return ptr && *ptr == n; // don't dereference a null pointer }

This code tests the value of ptr. If ptr is NULL, the second && operand is not evaluated, preventing a null pointer dereference. This behavior is a useful way to avoid unnecessary computing. Here, the is_file_ready predicate function will return true if the file is ready: is_file_ready() || prepare_file()

In this case, when the is_file_ready function returns true, the second || operand is not evaluated, as there is no need to prepare the file. This will avoid unnecessary computing, assuming the cost of determining whether the file is ready is less than the cost of preparing the file (and there’s a good chance the file is already prepared). Programmers should exercise caution if the second operand contains side effects, because it may not be apparent whether the side effects actually occur. For example, in the following code snippet, the value of i is incremented only when i >= 0: enum { max = 15 }; int i = 17; if ( (i >= 0) && ( (i++) (greater than), = (greater than or equal to). Each returns 1 if the specified relationship is true, and 0 if it is false. The result has type int, again, for historical reasons. Note that C does not interpret the expression a < b < c to mean that b is greater than a but less than c, as ordinary mathematics does. Instead, the expression is interpreted to mean (a < b) < c. In English, if a is less than b, the compiler should compare 1 to c; otherwise, it compares 0 to c. If this is your intent, you should include the parentheses to make this clear to any potential code reviewer. Some compilers such as GCC and Clang provide the –Wparentheses flag that diagnoses these problems. A test to determine whether b is greater than a but less than c can be written as (a < b) && (b < c). The equality and inequality operators have lower precedence than the relational operators—and assuming otherwise is a common mistake. This means that the expression a < b == c < d is evaluated the same as (a < b) == (c < d). In both cases, the comparisons a < b and c < d are evaluated first, and the resulting values (either 0 or 1) are compared for equality.

Expressions and Operators   77

We can use these operators to compare arithmetic types or pointers. When we compare two pointers, the result depends on the relative locations in the address space of the objects pointed to. If both pointers point to the same object, they are equal. The equality and inequality operators differ from the other relational operators. For example, you cannot use the other relational operators on two pointers to unrelated objects, because doing so makes no sense: int i, j; bool b1 = &i < &j; // undefined behavior bool b2 = &i == &j; // okay, but tautologically false

Compound Assignment Operators Compound assignment operators modify the current value of an object by performing an operation on it. The compound assignment operators are shown in Table 4-6. Table 4-6: Compound Assignment Operators

Operator

Description

+= -=

Assignment by sum and difference

*= /= %=

Assignment by product, quotient, and remainder

=

Assignment by bitwise left shift and right shift

&= ^= |=

Assignment by bitwise AND, XOR, and OR

A compound assignment of the form E1 op = E2 is equivalent to the simple assignment expression E1 = E1 op (E2), except that E1 is evaluated only once. Compound assignments are primarily used as a shorthand notation. There are no compound assignment operators for logical operators.

Comma Operator In C, we use commas in two distinct ways: as operators and as a way to separate items in a list (such as arguments to functions or lists of declarations). The comma (,) operator is a way to evaluate one expression before another. First, the left operand of a comma operator is evaluated as a void expression. There is a sequence point between the evaluation of the left operand and the evaluation of the right operand. Then, the right operand is evaluated after the left. The result of the comma operation has the type and value of the right operand—mostly because it is the last expression evaluated. You cannot use the comma operator in contexts in which a comma might separate items in a list. Instead, you would include a comma within

78   Chapter 4

a parenthesized expression, or within the second expression of a conditional operator. For example, the following function call has three parameters: f(a,u (t=3,v t+2),w c)

The first comma u separates the first and second arguments to the function. The second comma v is a comma operator. The assignment is evaluated first, followed by the addition. Because of the sequence point, the assignment is guaranteed to complete before the addition takes place. The result of the operation has the type (int) and value (5) of the right-hand operand. The third comma w separates the second and third arguments to the function.

Pointer Arithmetic Earlier in this chapter, we mentioned that it’s possible to use the additive operators (addition and subtraction) on either arithmetic values or object pointers. In this section, we discuss adding a pointer and an integer, subtracting two pointers, and subtracting an integer from a pointer. Adding or subtracting an expression that has integer type to or from a pointer returns a value with the type of the pointer operand. If the pointer operand points to an element of an array, then the result points to an e­ lement offset from the original element. If the resulting pointer is outside the bounds of the array, undefined behavior occurs. The difference of the array subscripts of the resulting and original array elements equals the integer expression: int arr[100]; int *arrp1 = arr[40]; int *arrp2 = arrp1 + 20; // arrp2 points to arr[60] printf("%td\n", arrp2-arrp1); // prints 20

C allows a pointer to be formed to each element of an array, including one past the last element of the array object (also referred to as the too-far pointer). While this might seem unusual or unnecessary, many early C programs incremented a pointer until it was equal to the too-far pointer, and the C Standards committee did not want to break all this code, which is also idiomatic in C++ iterators. Figure 4-3 illustrates forming the too-far pointer.

too-far

Figure 4-3: One past the last element of an array object

If both the pointer operand and the result point to elements of the same array object or the too-far pointer, the evaluation will not overflow; Expressions and Operators   79

otherwise, the behavior is undefined. To satisfy the too-far requirement, an implementation need only provide one extra byte (which can overlap another object in the program) just after the end of the object. C also allows objects to be treated as an array containing only a single element, allowing you to obtain a too-far pointer from a scalar. The too-far special case allows us to advance a pointer until it is equal to the too-far pointer, as in the following function: int m[2] = {1, 2}; int sum_m_elems(void) { int *pi; int j = 0; for (pi = &m[0]; pi < &m[2]; ++pi) j += *pi; return j; }

Here, the for statement (explained in detail in the next chapter) in the sum_m_elems function loops while pi is less than the address of the toofar pointer for the array m. The pointer pi is incremented at the end of each iteration of the loop until the too-far pointer is formed, causing the loop condition to subsequently evaluate to 0. When we subtract one pointer from another, both must point to elements of the same array object or the too-far element. This operation returns the difference of the subscripts of the two array elements. The type of the result is ptrdiff_t (a signed integer type). You should take care when subtracting pointers, because the range of ptrdiff_t may not be sufficient to represent the difference of pointers to elements of very large arrays of char. Pointer arithmetic is automatically scaled to work with array element sizes, rather than individual bytes.

Summary In this chapter, you learned how to use operators to write simple expressions that perform operations on various object types. Along the way, you learned about some core C concepts, such as lvalues, rvalues, value computations, and side effects, which determine how expressions are evaluated. You also learned how operator precedence, associativity, order of evaluation, sequencing, and interleaving can affect the total order in which a program is executed. In the next chapter, you’ll learn more about how to control the execution of your program by using selection, iteration, and jump statements.

80   Chapter 4

5 CONTROL FLOW

In this chapter, you’ll learn how to control the order in which individual statements are evaluated. We’ll start by going over expression statements and compound statements that define the work to be performed. We’ll then cover three kinds of statements that determine which code blocks are executed, and in what order: selection, iteration, and jump statements. Expression Statements An expression statement is an optional expression terminated by a semicolon. It is one of the most common kinds of statements and a basic unit of work. Listing 5-1 shows examples of expression statements.

a = 6; c = a + b; ; // null statement, does nothing ++count; Listing 5-1: Typical expression statements

The first statement consists of an expression that assigns a value to a. The second statement is an expression that assigns the sum of a and b to c. The third statement is a null statement, which you can use when the syntax of the language requires a statement but no expression needs to be evaluated. Null statements are commonly used as placeholders in iteration statements or as statements on which to place labels at the end of compound statements or functions. The fourth statement is an expression that increments the value of count. After each full expression has been evaluated, its value (if any) is discarded (including assignment expressions in which the assignment itself is a side effect of the operation) so that any useful results occur as the consequence of side effects (as discussed in Chapter 4). Three of the four expression statements in this example have side effects (the null statement does nothing). Once all side effects have completed, execution proceeds to the statement following the semicolon.

Compound Statements A compound statement, or block, is a list of zero or more statements, surrounded by braces. The statements in the block may be any kind of statement described throughout this chapter. Some of these statements may be declarations. (In early versions of C, declarations within the block had to precede all nondeclarations, but that restriction no longer applies.) Each statement in the block is executed in sequence, unless modified by a control statement. After the final statement has been evaluated, execution proceeds to just after the closing brace: { static int count = 0; c += a; ++count; }

This example declares a static variable of type int called count. The second line increases a variable c declared in an outer scope by the value stored in a. Finally, count is incremented to track how many times this block has been executed. Compound statements can be nested so that one compound statement fully encloses another. You may also have blocks with no statements at all (just the empty braces).

82   Chapter 5

CODE S T Y LE Competing coding styles disagree on when and where to place braces. If you’re modifying existing code, it would be wise to follow the style already in use for the project. Otherwise, take a look at styles you see in code written by experienced C programmers and choose one that seems clear. For example, some programmers line up the opening and closing braces to make it easy to find the mate for a given brace. Others follow the style used in The C Programming Language by Brian Kernighan and Dennis Ritchie (1988), wherein the opening brace is placed at the end of the preceding line, and the closing brace gets a line to itself. Once you have chosen a style, use it consistently.

Selection Statements Selection statements allow you to conditionally execute a substatement depending on the value of a controlling expression. The controlling expression determines which statements are executed based on a condition. This allows you to write code to produce different output based on different inputs. Selection statements include the if statement and the switch statement.

The if Statement The if statement allows a programmer to execute a substatement based on the value of a controlling expression of scalar type. There are two kinds of if statements. The first conditionally determines whether the substatement is executed: if (expression) substatement

In this case, substatement is executed if expression is not equal to 0. Only the single substatement of the if statement is conditionally executed, though it can be a compound statement. Listing 5-2 shows a division function that uses if statements. It divides a specified dividend by a specified divisor and returns the result in the object referenced by quotient. The function tests for both division by zero and signed integer overflow, and returns false in either case. bool safediv(int dividend, int divisor, int *quotient) { u if (!quotient) return false; v if ((divisor == 0) || ((dividend == INT_MIN) && (divisor == -1))) w return false; x *quotient = dividend / divisor; return true; } Listing 5-2: Safe division function Control Flow   83

The first line of this function u tests quotient to ensure that it’s not null. If it is null, the function returns false to indicate that it is unable to return a value. We cover return statements later in this chapter. The second line of the function v contains a more complex if statement. Its controlling expression tests whether the divisor is 0 or whether the division would result in signed integer overflow if unchecked. If the result of this expression is not equal to 0, the function returns false w to indicate that it is unable to produce a quotient. If the controlling expression of the if statement evaluates to 0, the function does not return, and the remaining statements x are executed. The second kind of if statement includes an else clause, which selects an alternative substatement to execute when the initial substatement is not selected: if (expression) substatement1 else substatement2

In this form, substatement1 is executed if expression is not equal to 0, and substatement2 is executed if the expression is equal to 0. It is always the case that one or the other of these substatements is executed, but never both. For either form of the if statement, the conditionally executed substatement may also be an if statement. A common use of this is the if...else ladder, shown in Listing 5-3. if (expr1) substatement1 else if (expr2) substatement2 else if (expr3) substatement3 else substatement4 Listing 5-3: if...else ladder

One (and only one) of the four statements in an if...else ladder will execute: • • •

substatement1 executes if expr1 does not equal 0.



substatement4 executes only if the proceeding conditions are all equal

substatement2 executes if expr1 equals 0 and if expr2 does not equal 0. substatement3 executes if both expr1 and expr2 equal 0 and expr3 does not

equal 0. to 0.

The example shown in Listing 5-4 uses an if..else ladder to print grades. 84   Chapter 5

void printgrade(unsigned int marks) { if (marks >= 90) { puts("YOUR GRADE : A"); } else if (marks >= 80) { puts("YOUR GRADE : B"); } else if (marks >= 70) { puts("YOUR GRADE : C"); } else { puts("YOUR GRADE : Failed"); } } Listing 5-4: Using an if...else ladder to print grades

In this if...else ladder, the printgrade function tests the value of the unsigned int parameter marks to determine whether it is greater than or equal to 90. If so, the function prints YOUR GRADE : A. Otherwise, it tests whether marks is greater than or equal to 80, and so forth down the if...else ladder. If marks is not greater than or equal to 70, the function prints YOUR GRADE : Failed. This example uses a coding style in which the closing brace is followed by the else clause on the same line. Only a single statement following the if statement is executed. For example, in the following code snippet, conditionally_executed_function is executed only if condition is not equal to 0, but unconditionally_executed  _function is always executed: if (condition) conditionally_executed_function(); unconditionally_executed_function(); // always executed

Attempting to add another conditionally executed function is a common source of errors: if (condition) conditionally_executed_function(); second_conditionally_executed_function(); // ???? unconditionally_executed_function(); // always executed

In this code snippet, second_conditionally_executed_function is unconditionally executed. The name and indented formatting are deceptive, because whitespace (in general) and indentation (in particular) are meaningless to the syntax. This code can be fixed by adding braces, delimiting a single compound statement or block, which is then executed as the single conditionally executed statement: if (condition) { conditionally_executed_function(); second_conditionally_executed_function(); // fixed it } unconditionally_executed_function(); // always executed

Control Flow   85

While the original code snippet was not incorrect, many coding guidelines recommend always including braces to avoid this kind of error: if (condition) { conditionally_executed_function(); } unconditionally_executed_function(); // always executed

My personal style is to omit the braces only when I can include the conditionally executed statement on the same line as the if statement: if (!quotient) return false;

This issue is less problematic when you let your IDE format your code for you, as it will not be fooled by the presence or absence of braces when formatting your code. Some compilers, such as GCC’s -Wmisleading-indentation, also check code indentation and issue warnings when it doesn’t correspond to the control flow.

The switch Statement The switch statement works just like the if...else ladder, except that the controlling expression must have an integer type. For example, the switch statement in Listing 5-5 performs the same function as the if...else ladder from Listing 5-4, provided that marks is an integer in the range of 0 to 100. If marks is greater than 109, it will result in a failed grade because the resulting quotient will be greater than 10 and will consequently be caught by the default case. switch (marks/10) { case 10: case 9: puts("YOUR GRADE break; case 8: puts("YOUR GRADE break; case 7: puts("YOUR GRADE break; default: puts("YOUR GRADE }

: A");

: B");

: C");

: Failed");

Listing 5-5: Using a switch statement to print out grades

The switch statement will cause control to jump to one of the three substatements, depending on the value of the controlling expression and the constant expressions in each case label. Following the jump, code is executed sequentially until the next control flow statement is reached. In our

86   Chapter 5

example, a jump to case 10 (which is empty) will flow through and execute the subsequent statements in case 9. This is necessary to the logic so that a perfect grade of 100 results in an A and not an F. You can terminate the execution of the switch, causing control to jump to the execution of the statement directly following the overall switch statement. We discuss break statements in more detail later in this chapter. Make sure you remember to include a break statement before the next case label. If omitted, the control flow will fall through to the next case in the switch statement—a common source of errors. Because the break statement is not required, omitting it doesn’t typically produce compiler diagnostics. GCC will issue a warning for fall-through if you use the -Wimplicit-fallthrough flag. The C2x standard will introduce a [[fallthrough]] attribute as a way for a programmer to specify that fall-through behavior is desirable, under the assumption that silent fall-through is an accidental omission of a break statement. Integer promotions are performed on the controlling expression. The constant expression in each case label is converted to the promoted type of the controlling expression. If a converted value matches that of the promoted controlling expression, control jumps to the statement following the matched case label. Otherwise, if there is a default label, control jumps to the labeled statement. If no converted case constant expression matches and there is no default label, no part of the switch body is executed. When switch statements are nested, a case or default label is accessible only within the closest enclosing switch statement. There are best practices regarding the use of switch statements. List­ ing 5-6 shows a not incorrect implementation of a switch statement that assigns interest rates to an account based on the account type. A fixed number of account types are offered by this bank, so the account types are represented by the AccountType enumeration. typedef enum { Savings, Checking, MoneyMarket } AccountType; void assignInterestRate(AccountType account) { double interest_rate; switch (account) { case Savings: interest_rate = 3.0; break; case Checking: interest_rate = 1.0; break; case MoneyMarket: interest_rate = 4.5; break; } printf("Interest rate = %g.\n", interest_rate); } Listing 5-6: switch statement without a default label

Control Flow   87

The assignInterestRate function defines a single parameter of the enumeration type assignInterestRate and switches on it to assign the appropriate interest rate associated with each account type. Nothing is wrong with the code as written, but it requires programmers to update the code in at least two separate places if they want to make any changes. Let’s say the bank introduces a new type of account: a certificate of deposit. The programmers update the AccountType enumeration as follows: typedef enum { Savings, Checking, MoneyMarket, CD } AccountType;

However, they neglect to modify the switch statement in the assign  InterestRate function. Therefore, interest_rate is not assigned, resulting in an uninitialized read when the function attempts to print this value. This is a common problem because the enumeration may be declared far from the switch statement, and the program may contain many similar switch statements that all reference an object of type AccountType in their controlling expression. Both Clang and GCC will help diagnose these problems at compilation time when you use the -Wswitch-enum flag. Alternatively, you can protect against such errors by including a default label in the switch: default: abort();

The abort function (declared in the header) causes abnormal program termination, making it easy to detect the error. You can improve the testability of this code by adding this default case to the switch statement, as shown in Listing 5-7. typedef enum { Savings, Checking, MoneyMarket, CD } AccountType; void assignInterestRate(AccountType account) { double interest_rate; switch (account) { case Savings: interest_rate = 3.0; break; case Checking: interest_rate = 1.0; break; case MoneyMarket: interest_rate = 4.5; break; case CD: interest_rate = 7.5; break; default: abort(); } printf("Interest rate = %g.\n", interest_rate); return; } Listing 5-7: switch statement with a default label

88   Chapter 5

The switch statement now includes a case for CD, and the default clause is unused. However, retaining the default clause is good practice, in case another account type is added in the future. Including a default clause does have the drawback of suppressing compiler warnings and not diagnosing the problem until runtime. Compiler warnings (if supported by your compiler) are therefore a better approach.

Iteration Statements Iteration statements cause substatements (or compound statements) to be executed zero or more times, subject to termination criteria. Iteration is an English word meaning “the repetition of a process.” Iteration statements are more informally and commonly referred to as loops. Loop is also an English word meaning “a process, the end of which is connected to the beginning.”

The while Statement The while statement causes the loop body to execute repeatedly until the controlling expression is equal to 0. The evaluation of the controlling expression occurs before each execution of the loop body. Consider the following example: void f(unsigned int x) { while (x > 0) { printf("%d\n," x); --x; } return; }

If x is not initially greater than 0, the while loop exits without executing the loop body. If x is greater than 0, its value is output and then decremented. Once the end of the loop is reached, the controlling expression is tested again. This pattern repeats until the expression evaluates to 0. Overall, this loop will count down from x to 1. A while loop is a simple entry-controlled loop that executes for as long as its entry condition is met. Listing 5-8 shows an implementation of the C Standard Library memset function. This function copies the value of val (converted to an unsigned char) into each of the first n characters of the object pointed to by dest. void *memset(void *dest, int val, size_t n) { unsigned char *ptr = (unsigned char*)dest; while (n-- > 0) *ptr++ = (unsigned char)val; return dest; } Listing 5-8: The C Standard Library memset function

Control Flow   89

The first line of the memset function converts dest to a pointer to an unsigned char and assigns the resulting value to the unsigned char pointer ptr. This lets us preserve the value of dest so it can be returned in the last line of the function. The remaining two lines of the function form a while loop that copies the value of val (converted to an unsigned char) into each of the first n characters of the object pointed to by dest. The controlling expression of the while loop tests that n-- > 0. The n argument is a loop counter that’s decremented on each iteration of the loop as a side effect of the evaluation of the controlling expression. The loop counter in this case monotonically decreases until a minimum value (0) is reached. The loop performs n repetitions, where n is less than or equal to the bound of the memory referenced by ptr. The pointer ptr designates a sequence of objects of type unsigned char, from ptr through ptr + n - 1. The value of val is converted to an unsigned char and written to each object in turn. If n is greater than the bound of the object referenced by ptr, the while loop will write to memory outside the bounds of this object. This is undefined behavior and a common security flaw, referred to as a buffer overflow, or overrun. Provided these preconditions are met, the while loop will terminate without undefined behavior. In the final iteration of the loop, the controlling expression n-- > 0 evaluates to 0, causing the loop to terminate. It is possible to write an infinite loop—a loop that runs forever. To avoid writing a while loop that inadvertently runs forever, be sure you initialize any objects referenced by the controlling expression before the start of the while loop. Also make sure that the controlling expression will change during the while loop’s execution in a manner that causes the loop to terminate after iterating an appropriate number of times.

The do...while Statement The do...while statement is similar to the while statement, except that the evaluation of the controlling expression takes place after each execution of the loop body, rather than before. As a result, the loop body is guaranteed to execute once before the condition is tested. The do...while iteration statement has the following syntax: do statement while ( expression );

In a do...while iteration, statement is always executed, after which expression is evaluated. If expression is not equal to 0, control returns to the top of the loop, and statement is executed again. Otherwise, execution passes to the statement following the loop. The do...while iteration statement is commonly used in I/O, where it makes sense to read from a stream before testing the state of the stream, as shown in Listing 5-9.

90   Chapter 5

#include //---snip--int count; float quant; char units[21], item[21]; do { count = fscanf(stdin, "%f%20s of %20s", &quant, units, item); fscanf(stdin,"%*[^\n]"); } while (!feof(stdin) && !ferror(stdin)); Listing 5-9: Repeatedly accepting a quantity, unit of measure, and an item name from stdin

This code inputs a floating-point quantity, a unit of measure (as a string), and an item name (also as a string) from the standard input stream stdin until the end-of-file indicator has been set or a read error has occurred. We’ll discuss I/O in detail in Chapter 8.

The for Statement The for statement might be the most C-like thing about C. The for statement repeatedly executes a statement and is typically used when the number of iterations is known before entering the loop. It has the following syntax: for (clause1; expression2; expression3) statement

The controlling expression expression2 is evaluated before each execution of the loop body, and expression3 is evaluated after each execution of the loop body. If clause1 is a declaration, the scope of any identifiers it declares is the remainder of the declaration and the entire loop, including the other two expressions. The purpose of clause1, expression2, and expression3 is apparent when we translate the for statement into an equivalent while loop, as shown in Figure 5-1.

for ( clause1 ; expression2 ; expression3 ) statement

clause1 while ( expression2 ) { statement (void) expression3 ; }

Figure 5-1: Translating a for loop into a while loop

Listing 5-10 shows a modified version of the memset implementation from Listing 5-8; we have replaced the while loop with a for loop. void *memset(void *dest, int val, size_t n) { unsigned char *ptr = (unsigned char *)dest; for (size_t i = 0;u i < n;v ++iw) { Control Flow   91

*(ptr + i) = val; } return dest; } Listing 5-10: Filling a character array by using a for loop

The for loop is popular among C programmers because it provides a convenient location for declaring and/or initializing the loop counter u, specifying the controlling expression for the loop v, and incrementing the loop counter w, all concisely on the same line. The for loop can also be somewhat misleading. Let’s take the example of a singly linked list in C that declares a node structure consisting of a data element and a pointer to the next node in the list. We also define a pointer p to the node structure: struct node { int data; struct node *next; }; struct node *p;

Using the definition of p, the following code snippet (used to deallocate the storage for a linked list) erroneously reads the value p after it has been freed: for (p = head; p != NULL; p = p->next) free(p);

Reading p after it has been freed is undefined behavior. If this loop were rewritten as a while loop, it would become apparent that the code reads p after it was freed: p = head; while (p != NULL) { free(p); p = p->next; }

The for loop can be confusing because it evaluates expression3 after the main body of the loop even though, lexically, it appears before the loop’s body. The correct way to perform this operation is to save the required pointer before freeing it, like this: for (p = head; p != NULL; p = q) { q = p->next; free(p); }

You can read more about dynamic memory management in Chapter 6. 92   Chapter 5

Jump Statements A jump statement unconditionally transfers control to another section of the same function when encountered. These are the lowest-level control flow statements and generally correspond closely to the underlying assembly ­language code.

The goto Statement Any statement may be preceded by a label, which is an identifier followed by a colon. A goto statement causes a jump to the statement prefixed by the named label in the enclosing function. The jump is unconditional, meaning it happens every time the goto statement is executed. Here is an example of a goto statement: /* executed statements */ goto location; /* skipped statements */ location: /* executed statements */

Execution continues until the goto statement is reached, at which point control jumps to the statement following the label location, where execution continues. Statements between the goto statement and the label are passed over. The goto statement has had a bad reputation since Edsger Dijkstra wrote a paper in 1968 titled “Go To Statement Considered Harmful” (1968). His criticism was that goto statements can result in spaghetti code if used haphazardly—code that has a complex and tangled control structure, resulting in a program flow that is conceptually twisted and tangled like a bowl of spaghetti. However, goto statements can also make code easier to read if used in a clear, consistent manner. One helpful way to use goto statements is to chain them together to release allocated resources (such as dynamic allocated memory or an open file) when an error occurs and you must leave a function. This scenario occurs when a program allocates multiple resources, each allocation can fail, and resources must be released to prevent leaking. If the first resource allocation fails, no cleanup is needed, because no resources have been allocated. However, if the second resource cannot be allocated, the first resource needs to be released. Similarly, if the third resource cannot be allocated, the second and first resources allocated need to be released, and so forth. This pattern results in duplicated cleanup code, and it can be error-prone because of the duplication and additional complexity. One solution is to use nested if statements, which can also become difficult to read if nested too deeply. Instead, we can use a goto chain as shown in Listing 5-11 to release resources. int do_something(void) { FILE *file1, *file2; object_t *obj; Control Flow   93

int ret_val = 0; // Initially assume a successful return value file1 = fopen("a_file", "w"); if (file1 == NULL) { ret_val = -1; goto FAIL_FILE1; } file2 = fopen("another_file", "w"); if (file2 == NULL) { ret_val = -1; goto FAIL_FILE2; } obj = malloc(sizeof(object_t)); if (obj == NULL) { ret_val = -1; goto FAIL_OBJ; } // Operate on allocated resources // Clean up everything free(obj); FAIL_OBJ: // Otherwise, close only the resources we opened fclose(file2); FAIL_FILE2: fclose(file1); FAIL_FILE1: return ret_val; } Listing 5-11: Using a goto chain to release resources

The code follows a simple pattern: resources are allocated in a certain order, operated upon, and then released in reverse (last in, first out) order. If an error occurs while allocating a resource, the code uses a goto to jump to the appropriate location in cleanup code and releases only those resources that have been allocated. Used in a structured manner such as this, goto statements can make code easier to read. A real-world example is the copy_process function from kernel/fork.c of the Linux kernel, which uses 17 goto labels to perform cleanup code when an internal function fails.

The continue Statement You can use a continue statement inside a loop to jump to the end of the loop body, skipping the execution of the remaining statements inside the loop body for the current iteration. For example, the continue statement is equivalent to goto END_LOOP_BODY; in each of the loops shown in Listing 5-12.

94   Chapter 5

while (/* … */) { //---snip-- continue; //---snip--END_LOOP_BODY: ; }

do { //---snip-- continue; //---snip--END_LOOP_BODY: ; } while (/* … */);

for (/* … */) { //---snip-- continue; //---snip--END_LOOP_BODY: ; }

Listing 5-12: Using the continue statement

The continue statement is frequently used in conjunction with a conditional statement so that processing may continue with the subsequent loop iteration after the objective of the current loop iteration has been achieved.

The break Statement A break statement terminates execution of a switch or iteration statement. We used break within a switch statement earlier in this chapter. Within a loop, a break statement causes the loop to terminate and the program execution to resume at the statement following the loop. For example, the for loop in the following example will exit only when the uppercase or lowercase Q key is pressed on the keyboard: #include int main(void) { char c; for(;;) { puts("Press any key, Q to quit: "); c = toupper(getchar()); if (c == 'Q') break; } } // Loop exits only when the uppercase or lowercase 'Q' is pressed

We typically use break statements to discontinue the execution of the loop when the work it was performing has been completed. For example, the break statement in Listing 5-13 exits the loop after it finds the specified key in an array. Assuming that key is unique in arr, the find_element function would behave the same without the break statement, but could negatively affect performance by executing many more instructions, depending on the length of the array and the point at which key is discovered. size_t find_element(size_t len, int arr[len], int key) { size_t pos = (size_t)-1; // traverse arr and search for key for (size_t i = 0; i < len; ++i) { if (arr[i] == key) { pos = i; break; // terminate loop } }

Control Flow   95

return pos; } Listing 5-13: Breaking out of a loop

Because continue and break bypass part of a loop body, you should use these statements carefully: the code following these statements is not executed.

The return Statement A return statement terminates execution of the current function and returns control to its caller. You’ve already seen many examples of return statements in this book. A function may have 0 or more return statements. A return statement can simply return, or it can return an expression. Within a void function (a function that doesn’t return a value), the return statement should simply return. When a function returns a value, the return statement should return an expression that produces a value of the return type. If a return statement with an expression is executed, the value of the expression is returned to the caller as the value of the function call expression: int sum(int x, int y, int z) { return x + y + z; }

This simple function sums its parameters and returns the sum. The return expression x + y + z produces a value of type int, which matches the return type of the function. If this expression produced a different type, it would be implicitly converted to an object having the return type of the function. The return expression can also be as simple as returning 0 or 1. The function result may then be used in an expression or assigned to a variable. Be aware that if control reaches the closing brace of a non-void function (a function declared to return a value) without evaluating a return statement with an expression, then using the return value of the function call is undefined behavior. For example, the following function fails to return a value when a is nonnegative, because the condition a < 0 is false: int absolute_value(int a) { if (a < 0) { return -a; } }

We can easily repair this defect by providing a return value when a is nonnegative, as shown in Listing 5-14.

96   Chapter 5

int absolute_value(int a) { if (a < 0) { return -a; } return a; } Listing 5-14: The absolute_value function returns a value along all paths.

However, this code still has a bug if two’s-complement representation is used (see Chapter 3). Identifying this bug is left as an exercise for you.

Exercises Try these coding exercises on your own: 1. Modify the function from Listing 5-11 to make it clear to the caller which file could not be opened. 2. Change the find_element function from Listing 5-13 to return the position of the key in a. Don’t forget to return an error indication if the key is not found. 3. Fix the remaining bug in the absolute_value function in Listing 5-14.

Summary In this chapter, you learned about control flow statements: • • •

Selection statements, such as if and switch, allow you to select from a set of statements depending on the value of a controlling expression. Iteration statements repeatedly execute a loop body until a controlling expression equals 0. Jump statements unconditionally transfer control to a new location. In the next chapter, you’ll learn about characters and strings.

Control Flow   97

6 D Y N A M I C A L LY A LLOC AT E D ME MORY

In Chapter 2, you learned that every object has a storage duration that determines its lifetime, and that C defines four storage durations: static, thread, automatic, and allocated. In this chapter, you’ll learn about dynamically allocated memory, which is allocated from the heap at runtime. Dynamically allocated memory is useful when the exact storage requirements for a program are unknown before runtime. We’ll first describe the differences between allocated, static, and automatic storage duration. We’ll skip thread storage allocation as this involves parallel execution, which we don’t cover here. We’ll then explore the functions you can use to allocate and deallocate dynamic memory, common memory allocation errors, and strategies for avoiding them. The terms memory and storage are used interchangeably in this chapter, in similar fashion to the way they’re used in practice.

Storage Duration Objects occupy storage, which might be random-access memory (RAM), read-only memory (ROM), or registers. Storage of allocated duration has significantly different properties than storage of either automatic or static storage duration. First, we’ll review automatic and static storage duration— initially described in Chapter 2. Objects of automatic storage duration are declared within a block or as a function parameter. The lifetime of these objects begins when the block in which they’re declared begins execution, and ends when execution of the block ends. If the block is entered recursively, a new object is created each time, each with its own storage. Objects declared in the file scope have static storage duration. The lifetime of these objects is the entire execution of the program, and their stored value is initialized prior to program startup. You can also declare a variable within a block scope to have static storage duration by using the storage-class specifier static.

The Heap and Memory Managers Dynamically allocated memory has allocated storage duration. The lifetime of an allocated object extends from the allocation until the deallocation. Dynamically allocated memory is allocated from the heap, which is simply one or more large, subdividable blocks of memory that are managed by the memory manager. Memory managers are libraries that manage the heap for you by providing implementations of the standard memory management functions described in the following sections. A memory manager runs as part of the client process. The memory manager will request one or more blocks of memory from the operating system and then allocate this memory to the client process when it invokes a memory allocation function. Memory managers manage unallocated and deallocated memory only. Once memory has been allocated, the caller manages the memory until it’s returned. It’s the caller’s responsibility to ensure that the memory is deallocated, although most implementations will reclaim dynamically allocated memory when the program terminates.

ME MORY M A N AGE R IMPLE ME NTATIONS Memory managers typically implement a variant of a dynamic storage allocation algorithm described by Donald Knuth (1997). This algorithm uses boundary tags—size fields that appear before and after the block of memory returned to the programmer. This size information allows all memory blocks to be traversed from any known block in either direction, allowing the memory manager to coalesce two bordering unused blocks into a larger block to minimize fragmentation.

100   Chapter 6

Fragmentation occurs when memory is allocated and deallocated, resulting in many small blocks of memory but no remaining large blocks. As a result, larger allocations can fail even though the total amount of free memory is sufficient for the allocation. Memory allocated for the client process and memory allocated for internal use within the memory manager is all within the addressable memory space of the client process.

When to Use Dynamically Allocated Memory Dynamically allocated memory is used when the exact storage requirements for a program are unknown before runtime. Dynamically allocated memory is less efficient than statically allocated memory, because the memory manager needs to find appropriately sized blocks of memory in the runtime heap, and then the caller must explicitly free those blocks when no longer needed, all of which requires additional processing. By default, you should declare objects with either automatic or static storage duration for objects whose sizes are known at compilation time. Memory leaks occur when dynamically allocated memory that’s no longer needed isn’t returned to the memory manager. If these memory leaks are severe, the memory manager will eventually be unable to satisfy new requests for storage. Dynamically allocated memory also requires additional processing for housekeeping operations such as defragmentation (the consolidation of adjacent free blocks), and the memory manager often uses extra storage for control structures to facilitate these processes. You would typically use dynamically allocated memory when the size of the storage isn’t known at compile time or the number of objects isn’t known until runtime. For example, you might use dynamically allocated memory to read a table from a file at runtime, especially if you don’t know the number of rows in the table at compile time. Similarly, you might use dynamically allocated memory to create linked lists, hash tables, binary trees, or other data structures for which the number of data elements held in each container is unknown at compile time.

Memory Management Functions The C Standard Library defines memory management functions for allocating and deallocating dynamic memory. These functions include malloc, aligned_alloc, calloc, and realloc. You can call the free function to deallocate memory. The OpenBSD reallocarray function isn’t defined by the C Standard Library but can be useful for memory allocation.

Dynamically Allocated Memory   101

The malloc Function The malloc function allocates space for an object of a specified size whose initial value is indeterminate. In Listing 6-1, we call the malloc function to dynamically allocate storage for an object the size of struct widget. #include typedef struct { char c[10]; int i; double d; } widget; u widget *p = malloc(sizeof(widget)); v if (p == NULL) { // Handle allocation error } // Continue processing Listing 6-1: Using the malloc function to allocate storage for a widget

All memory allocation functions accept an argument of type size_t that specifies the number of bytes of memory to be allocated u. For portability, we use the sizeof operator when calculating the size of objects, as the size of objects of various types, such as int and long, may differ among implementations. The malloc function returns either a null pointer to indicate an error, or a pointer to the allocated space. Therefore, we check whether malloc returns a null pointer v and appropriately handle the error. After the function successfully returns the allocated storage, we can store to members of the widget structure through the pointer p. For e­ xample, p->i will access the int member of widget, while p->d will access the double member. Allocating Memory Without Declaring a Type You can store the return value from malloc as a void pointer to avoid declaring a type for the referenced object: void *p = malloc(size);

Alternately, you can use a char pointer, which was the convention before the void type was introduced to C: char *p = malloc(size);

In either case, the object referenced by p has no type until an object is copied into this storage. Once this occurs, the object has the effective type of the last object copied into this storage, which imprints the type onto the allocated object. In the following example, the storage referenced by p has an effective type of widget following the call to memcpy. 102   Chapter 6

widget w = {"abc", 9, 3.2}; memcpy(p, &w, sizeof(widget)); // coerced to void * pointers printf("p.i = %d.\n", p->i);

Because any type of object can be stored in allocated memory, we can assign the pointers returned by all allocation functions, including malloc, to point to any type of object. For example, if an implementation has objects with 1-, 2-, 4-, 8-, and 16-byte alignments, and 16 or more bytes of storage are allocated, a pointer will be returned whose alignment is a multiple of 16. Casting the Pointer to the Type of the Declared Object Even expert C programmers disagree about whether to cast the pointer returned by malloc to a pointer to the type of the declared object. The following assignment statement casts this pointer as a pointer to widget: widget *p = (widget *)malloc(sizeof(widget));

Strictly speaking, this cast is unnecessary. C will allow you to implicitly convert a pointer to void (the type returned by malloc) to a pointer to any type of object for which the resulting pointer is correctly aligned. (Otherwise, the behavior is undefined.) Casting the result of malloc to the intended pointer type enables the compiler to catch inadvertent pointer conversions, as well as differences between the size of the allocation and the size of the pointed-to type in the cast expression. Examples in this book typically use a cast, but either style is acceptable. See CERT C rule MEM02-C (Immediately cast the result of a memory allocation function call into a pointer to the allocated type) for more on this topic. Reading Uninitialized Memory The contents of memory returned from malloc are uninitialized, which means it contains indeterminate values. Reading uninitialized memory is never a good idea, and you should think of it as undefined behavior—if you’d like to know more, I wrote an in-depth article on uninitialized reads (Seacord 2017). The malloc function does not initialize the returned memory, because the expectation is that you will overwrite this memory anyway. Even so, beginners commonly make the mistake of assuming that the memory returned by malloc contains zeros. The program shown in Listing 6-2 makes this exact error. #include #include #include int main(void) { char *str = (char *)malloc(16); if (str) { strncpy(str, "123456789abcdef", 15); Dynamically Allocated Memory   103

printf("str = %s.\n", str); free(str); return EXIT_SUCCESS; } return EXIT_FAILURE; } Listing 6-2: Initialization error

This program allocates 16 bytes of memory by calling malloc, and then uses strncpy to copy the first 15 bytes of a string into the allocated memory. The programmer attempts to create a properly null-terminated string by copying one less byte than the size of the allocated memory. In doing so, the programmer assumes that the allocated storage already contains a 0 value to serve as the null byte. However, the storage could easily contain nonzero values, in which case the string wouldn’t be properly null-­ terminated, and the call to printf would result in undefined behavior. A common solution is to write a null character into the last byte of the allocated storage, as follows: strncpy(str, "123456789abcdef", 15); u str[15] = '\0';

If the source string is less than 15 bytes, the null termination character will be copied, and the assignment at u is unnecessary. If the source string is 15 bytes or longer, adding this assignment will ensure that the string is properly null-terminated.

The aligned_alloc Function The aligned_alloc function is similar to the malloc function, except that it requires you to define an alignment as well as a size for the allocated object. The function has the following signature, where size specifies the object’s size, and alignment specifies its alignment: void *aligned_alloc(size_t alignment, size_t size);

C11 introduced the aligned_alloc function because some hardware has stricter-than-normal memory alignment requirements. Although C requires the dynamically allocated memory from malloc to be sufficiently aligned for all standard types, including arrays and structures, you might occasionally need to override the compiler’s default choices. Typically, we use the aligned_alloc function to request stricter alignment than the default (in other words, a larger power of two). If the value of alignment is not a valid alignment supported by the implementation, the function fails by returning a null pointer. See Chapter 2 for more information on alignment.

104   Chapter 6

The calloc Function The calloc function allocates storage for an array of nmemb objects, each of whose size is size bytes. It has the following signature: void *calloc(size_t nmemb, size_t size);

This function initializes the storage to all zero-valued bytes. This zero value might not be the same one used to represent floating-point zero or null-pointer constants. You can also use the calloc function to allocate storage for a single object, which can be thought of as an array of one element. Internally, the calloc function works by multiplying nmemb by size to determine the required number of bytes to allocate. Historically, some c­ alloc implementations failed to validate that these values wouldn’t overflow when multiplied. Modern implementations of calloc perform this overflow check and return a null pointer if the product cannot be represented in size_t.

The realloc Function The realloc function increases or decreases the size of previously allocated storage. It takes a pointer to some memory allocated by an earlier call to aligned_alloc, malloc, calloc, or realloc (or a null pointer) and a size, and has the following signature: void *realloc(void *ptr, size_t size);

You can use the realloc function to grow or (less commonly) shrink the size of an array. Avoiding Memory Leaks To avoid introducing bugs when you use realloc, you should understand how the function is (conceptually) implemented. The realloc function typically calls the malloc function to allocate new storage, and then copies the contents of the old storage to the new storage up to the minimum of the old and new sizes. If the newly allocated storage is larger than the old contents, realloc leaves the additional storage uninitialized. If realloc succeeds in allocating the new object, it calls free to deallocate the old object. If the allocation fails, the realloc function retains the old object data at the same address and returns a null pointer. A call to realloc can fail, for example, when insufficient memory is available to allocate the requested number of bytes. The following use of realloc contains an error: size += 50; if ((p = realloc(p, size)) == NULL) return NULL;

Dynamically Allocated Memory   105

In this example, size is incremented by 50 before calling realloc to increase the size of the storage referenced by p. If the call to realloc fails, p is assigned the value NULL, but realloc doesn’t deallocate the storage referenced by p, resulting in this memory being leaked. Listing 6-3 demonstrates the correct use of the realloc function. void *p2; void *p = malloc(100); //---snip--if ((nsize == 0) || (p2 = realloc(p, nsize)) == NULL) { free(p); return NULL; } p = p2; Listing 6-3: Correct use of the realloc function

This code snippet declares two variables, p and p2. The variable p refers to the dynamically allocated memory returned by malloc, and p2 starts out uninitialized. Eventually, this memory is resized, which we accomplish by calling the realloc function with the pointer p and the new size nsize. The return value from realloc is assigned to p2 to avoid overwriting the pointer stored in p. If realloc returns a null pointer, the memory referenced by p is freed, and the function returns a null pointer. If realloc succeeds and returns a pointer to an nsize allocation, p is assigned the pointer to the newly reallocated storage, and execution continues. This code also includes a test for a zero-byte allocation. You should avoid passing the realloc function a value of 0 as the size argument, as this is effectively undefined behavior (and is actually undefined behavior in C2x). If the following call to the realloc function does not return a null pointer, the address stored in p is invalid and can no longer be read: newp = realloc(p, ...);

In particular, the following test is not allowed: if (newp != p) { // update pointers to reallocated memory }

Any pointers that reference the memory previously pointed to by p must be updated to reference the memory pointed to by newp after the call to realloc regardless of whether realloc kept the same address for the storage. One solution to this problem is to go through an extra indirection, sometimes called a handle. If all uses of the reallocated pointer are indirect, they will all be updated when that pointer is reassigned.

106   Chapter 6

Calling realloc with a Null Pointer Calling realloc with a null pointer is equivalent to calling malloc. Provided newsize isn’t equal to 0, the following code if (p == NULL) newp = malloc(newsize); else newp = realloc(p, newsize);

can be replaced with newp = realloc(p, newsize);

The first, longer version of this code calls malloc the first time storage is allocated, and realloc to adjust the size later as required. Because calling realloc with a null pointer is equivalent to calling malloc, the second version concisely accomplishes the same thing.

The reallocarray Function The OpenBSD reallocarray function can reallocate storage for an array, but also provides overflow checking for array size calculations. This saves you from having to perform these checks. The reallocarray function has the following signature: void *reallocarray(void *ptr, size_t nmemb, size_t size);

The reallocarray function allocates storage for nmemb members of size size and checks for integer overflow in the calculation nmemb * size. Other platforms, including the GNU C Library (libc), have adopted this function, and it has been proposed for inclusion in the next revision of the POSIX standard. The reallocarray function doesn’t zero out the allocated storage. As we have seen in previous chapters, integer overflow is a serious problem that can result in buffer overflows and other security vulnerabilities. In the following code, for example, the expression num * size might overflow before being passed as the size argument in the following call to realloc: if ((newp = realloc(p, num * size)) == NULL) { //---snip---

The reallocarray function is useful when two values are multiplied to determine the size of the allocation: if ((newp = reallocarray(p, num, size)) == NULL) { //---snip---

This call to the reallocarray function will fail and return a null pointer if num * size would otherwise overflow.

Dynamically Allocated Memory   107

The free Function You should deallocate dynamically allocated memory when it’s no longer needed by calling the free function. Deallocating memory is important because it allows that memory to be reused, reducing the chances that you’ll exhaust the available memory, and often providing more efficient use of the heap. We deallocate memory by passing a pointer to that memory to the free function, which has the following signature: void free(void *ptr);

The ptr value must have been returned by a previous call to aligned_alloc, malloc, calloc, or realloc. CERT C rule MEM34-C (Only free memory allocated dynamically) discusses what happens when the value is not returned. Memory is a limited resource and so must be reclaimed. If we call free with a null-pointer argument, nothing happens, and the free function simply returns: char *ptr = NULL; free(ptr);

Avoiding Double-Free Vulnerabilities If you call the free function on the same pointer more than once, undefined behavior occurs. These defects can result in a security flaw known as a d­ ouble-free vulnerability. One possible consequence is that they can be exploited to execute arbitrary code with the permissions of the vulnerable process. The full effects of double-free vulnerabilities are beyond the scope of this book, but I discuss them in detail in Secure Coding in C and C++ (Seacord 2013). Double-free vulnerabilities are especially common in error-handling code, as programmers attempt to free allocated resources. Another common error is to access memory that has already been freed. This type of error frequently goes undiagnosed, because the code might appear to work, but fails in unexpected ways away from the actual error. In Listing 6-4, taken from an actual application, the argument to close is invalid, because the storage formerly pointed to by dirp has been reclaimed by the second call to free: #include #include #include int closedir(DIR *dirp) { free(dirp->d_buf); free(dirp); return close(dirp->d_fd); // dirp has already been freed } Listing 6-4: Accessing already freed memory

108   Chapter 6

We refer to pointers to already freed memory as dangling pointers. Dangling pointers are a potential source of errors (like a banana peel on the floor) because they can be used to write to memory that has already been freed or passed to the free function, resulting in double-free vulnerabilities. See CERT C rule MEM30-C (Do not access freed memory) for more information on these topics. Setting the Pointer to Null To limit the opportunity for defects involving dangling pointers, set the pointer to NULL after completing a call to free: char *ptr = malloc(16); //---snip--free(ptr); ptr = NULL;

Any future attempt to dereference the pointer will usually result in a crash (increasing the likelihood that the error is detected during implementation and testing). If the pointer is set to NULL, the memory can be freed multiple times without consequence. Unfortunately, the free function cannot set the pointer to NULL itself, because it’s passed a copy of the pointer, not the actual pointer.

Memory States Dynamically allocated memory can exist in one of three states shown in Figure 6-1: unallocated and uninitialized within the memory manager, allocated but uninitialized, and allocated and initialized. Calls to the ­malloc and free functions, as well as writing the memory, cause the memory to transition from one state to another. Unallocated and uninitialized: invalid to read, write, or free malloc free

Allocated but uninitialized: valid to write or free, but invalid to read

free

Allocated and initialized: valid to read and write or free

write

Figure 6-1: Memory states Dynamically Allocated Memory   109

Different operations are valid depending on the state of the memory. Avoid any operation on memory that is not shown as valid or is explicitly listed as invalid. This applies to each byte of memory, as bytes that have been initialized can be read, but bytes that have not been initialized must not be read.

Flexible Array Members Allocating storage for a structure that contains an array has always been a little tricky in C. There’s no problem if the array has a fixed number of ­elements, as the size of the structure can easily be determined. Developers, however, frequently need to declare an array with a flexible size, and previously, C offered no easy way to do this. Flexible array members let you declare and allocate storage for a structure with any number of fixed members, where the last member is an array of unknown size. Starting with C99, the last member of a struct with more than one member can have incomplete array type, which means that the array has an unknown size. This enables you to wait until runtime to specify the array size. A flexible array member allows you to access a variable-length object. For example, Listing 6-5 shows the use of a flexible array member data in widget. We dynamically allocate storage for the object by calling the m­ alloc function. #include typedef struct { size_t num; u int data[]; } widget; void *func(size_t array_size) { v widget *p = (widget *)malloc(sizeof(widget) + sizeof(int) * array_size); if (p == NULL) { return NULL; } p->num = array_size; for (size_t i = 0; i < p->num; ++i) { w p->data[i] = 17; } } Listing 6-5: Flexible array members

We first declare a struct whose last member, the data array u, is an incomplete type (with no specified size). We then allocate storage for the entire struct v. When computing the size of a struct containing a flexible array member using the sizeof operator, the flexible array member is ignored. Therefore, we must explicitly include an appropriate size for the flexible array member when allocating storage. To accomplish this, we allocate additional bytes for the array by multiplying the number of e­ lements 110   Chapter 6

in the array (array_size) by the size of each element (sizeof(int)). This program assumes that the value of array_size is such that when multiplied by sizeof(int) an overflow will not occur. We can access this storage by using a . or -> operator w, as if the storage had been allocated as data[array_size]. See CERT C rule MEM33-C (Allocate and copy structures containing a flexible array member dynamically) for more information on allocating and copying structures containing flexible array members. Prior to C99, multiple compilers supported a similar “struct hack” using a variety of syntaxes. CERT C rule DCL38-C (Use the correct syntax when declaring a flexible array member) is a reminder to use the syntax specified in C99 and later versions of the C Standard.

Other Dynamically Allocated Storage Beyond the memory management functions that allow you to allocate memory from the heap, C has language and library features that allow storage to be dynamically allocated. This storage is typically allocated in the stack frame of the caller (the C Standard does not define a stack, but it is a common implementation feature). A stack is a last-in-first-out (LIFO) data structure that supports nested invocation of functions at runtime. Each function invocation creates a stack frame in which local variables (of automatic storage duration) and other data specific to that invocation of the function can be stored.

The alloca Function For performance reasons, the alloca function, supported by some implementations, allows dynamic allocation at runtime from the stack rather than the heap. This memory is automatically released when the function that called alloca returns. The alloca function is an intrinsic (or built-in) function, which means that its implementation is handled specially by the compiler. This allows the compiler to substitute a sequence of automatically generated instructions for the original function call. For example, on the x86 architecture, the compiler substitutes a call to alloca with a single instruction to adjust the stack pointer to accommodate the additional storage. The alloca function originated in an early version of the Unix operating system from Bell Laboratories but is not defined by the C Standard Library or POSIX. Listing 6-6 shows an example function called printerr that uses the alloca function to allocate storage for an error string before printing it out to stderr. void printerr(errno_t errnum) { rsize_t size = strerrorlen_s(errnum) + 1; char *msg = (char *)alloca(size); if (strerror_s(msg, size, errnum) != 0) { fputs(msg, stderr); } else { Dynamically Allocated Memory   111

fputs("unknown error", stderr); } } Listing 6-6: The printerr function

The printerr function takes a single argument, errnum, of type errno_t. In the first line of the function, we call the strerrorlen_s function to determine the length of the error string associated with this particular error number. Once we know the size of the array that we need to allocate to hold the error string, we can call the alloca function to efficiently allocate storage for the array. We then retrieve the error string by calling the strerror_s function and store the result in the newly allocated storage referenced by msg. Assuming the strerror_s function succeeds, we output the error message; otherwise, we output unknown error. This printerr function is written to demonstrate the use of alloca and is more complicated than it actually needs to be. The alloca function can be tricky to use. First, the call to alloca can make allocations that exceed the bounds of the stack. However, the alloca function does not return a null pointer value, so there is no way to check for the error. For this reason, it’s critically important to avoid using alloca with large or unbounded allocations. The call to strerrorlen_s in this example should return a reasonable allocation size. A further problem with the alloca function is that programmers may become confused by having to free calls to malloc but not to alloca. Calling free on a pointer not obtained by calling aligned_alloc, calloc, realloc, or malloc is a serious error. Compilers also tend to avoid inlining functions that call alloca. For these reasons, the use of alloca is discouraged. The GCC compiler provides a -Walloca flag that diagnoses all calls to the alloca function, and a -Walloca-larger-than=size flag that diagnoses any call to the alloca function when the requested memory is more than size.

Variable-Length Arrays Variable-length arrays (VLAs), introduced in C99, are arrays that you can declare by using a variable to specify the array dimensions, allowing you to specify their size at runtime. The size of the array cannot be modified after you create it. VLAs are useful when you don’t know the number of elements in the array until runtime. All VLA declarations must be at either block scope or function prototype scope. The next sections show examples of each. Block Scope The following function func declares the variable-length array vla of size size as an automatic variable with block scope. void func(size_t size) { int vla[size]; //---snip--}

112   Chapter 6

The array is allocated in the stack frame and freed when the current frame exits—similar to the alloca function. Listing 6-7 replaces the call to alloca in the printerr function from Listing 6-6 with a VLA. The change modifies just a single line of code (shown in bold). void print_error(int errnum) { size_t size = strerrorlen_s(errnum) + 1; char msg[size]; if (strerror_s(msg, size, errnum) != 0) { fputs(msg, stderr); } else { fputs("unknown error", stderr); } } Listing 6-7: The print_error function rewritten to use a VLA

The main advantage of using VLAs instead of the alloca function is that the syntax matches the programmer’s model of how arrays with automatic storage duration work—that is, there’s no need to explicitly free the storage. VLAs share some of the problems of the alloca function, in that they can attempt to make allocations that exceed the bounds of the stack. Unfortunately, there’s no portable way to determine the remaining stack space to detect such an error. Also, the calculation of the array’s size could overflow when the size you provide is multiplied by the size of each element. For these reasons, it’s important to validate the size of the array before declaring it to avoid overly large or incorrectly sized allocations. This can be especially important in functions that are called recursively, because a complete new set of automatic variables for the functions (including these arrays) will be created for each recursion. You should determine whether you have sufficient stack space in the worst-case scenario (maximum-sized allocations with deep recursions). On some implementations, it’s also possible to pass a negative size to the VLA, so make sure your size is represented as a size_t or other unsigned type. See CERT C rule ARR32-C (Ensure size arguments for variable-length arrays are in a valid range) for more information. For GCC, you can use the -Wvla-larger-than=size flag to diagnose definitions of VLAs that either exceed the specified size or whose bound is not sufficiently constrained. Finally, another interesting and possibly unexpected behavior occurs when calling sizeof on a VLA. The compiler usually performs the sizeof operation at compile time. However, if the expression changes the size of the array, it will be evaluated at runtime, including any side effects. The same is true of typedef, as shown by the program in Listing 6-8. #include #include int main(void) { Dynamically Allocated Memory   113

size_t size = 12; printf("%zu\n", size); // prints 12 (void)sizeof(int[size++]); printf("%zu\n", size); // prints 13 typedef int foo[size++]; printf("%zu\n", size); // prints 14 } Listing 6-8: Unexpected side effects

In this simple test program, we declare a variable size of type size_t and initialize it to 12. We then call the sizeof operator with int[size++] as the argument. Because this expression changes the size of the array, size is incremented and is now equal to 13. The typedef similarly increments the value of size to 14. Function Prototype Scope You can also declare VLAs as function parameters. Remember from Chapter 2 that, when used in an expression, an array is converted to a pointer to the first element of the array. This means that we must add an explicit parameter to specify the size of the array—for example, the n parameter in the signature for memset: void *memset(void *s, int c, size_t n);

When you call such a function, n should accurately represent the size of the array referenced by s. Undefined behavior results if this size is larger than the array. When declaring a function to take a VLA as an argument that specifies a size, we must declare the size of the array before referencing the size in the array declaration. We could, for example, modify the signature for the memset function as follows to take a VLA: void *memset_vla(size_t n, char s[n], int c);

Here, we’ve changed the order of the parameters so that the variable n of type size_t is declared before we use it in the array declaration. The array argument s is still demoted to a pointer, and no storage is allocated as

a result of this declaration. When calling this function, you are responsible for declaring the actual storage for the array referenced by s and ensuring that n is a valid size for it. VLAs can generalize your functions, making them more useful. For example, the matrix_sum function sums all the values in a two-dimensional array. The following version of this function accepts a matrix with a fixed column size: int matrix_sum(size_t rows, int m[][4]);

114   Chapter 6

When passing a multidimensional array to a function, the size information for the initial dimension of the array is lost, but it still needs to be passed in as an argument. This information is provided by the rows parameter in this example. You can call this function to sum the values of any matrix with exactly four columns, as shown in Listing 6-9. int main(void) { int m1[5][4]; int m2[100][4]; int m3[2][4]; printf("%d.\n", matrix_sum(5, m1)); printf("%d.\n", matrix_sum(100, m2)); printf("%d.\n", matrix_sum(2, m3)); } Listing 6-9: Summing matrices with four columns

This is all well and good until you need to sum the values of a matrix that doesn’t have four columns. For example, changing m3 to have five columns would result in a warning such as this: warning: incompatible pointer types passing 'int [2][5]' to parameter of type 'int (*)[4]'

To handle this case, you’d have to write a new function with a signature that matches the new dimensions of the multidimensional array. The problem with this approach, then, is that it fails to generalize sufficiently. Instead of doing that, we can rewrite the matrix_sum function to use a VLA, as shown in Listing 6-10. This change allows us to call matrix_sum with matrices of any dimension. int matrix_sum(size_t rows, size_t cols, int m[rows][cols]) { int total = 0; for (size_t r = 0; r < rows; r++) for (size_t c = 0; c < cols; c++) total += m[r][c]; return total; } Listing 6-10: Using a VLA as a function parameter

Again, no storage is allocated by either the function declaration or the function definition. You need to allocate the storage for the matrix separately, and its dimensions must match those passed to the function as the rows and cols arguments. Failing to do so can result in undefined behavior.

Debugging Allocated Storage Problems As noted earlier in this chapter, improper memory management can lead to errors like leaking memory, reading from or writing to freed memory, and freeing memory more than once. One way to avoid some of these problems Dynamically Allocated Memory   115

is to set pointers to NULL after calling free, as we’ve already discussed. Another strategy is to keep your dynamic memory management as simple as possible. For example, you should allocate and free memory in the same module, at the same level of abstraction, rather than freeing memory in subroutines, which leads to confusion about if, when, and where memory was freed. A third option is to use dynamic analysis tools, which detect and report memory errors. These tools, as well as general approaches to debugging, testing, and analysis, are discussed in Chapter 11. In this section, we’ll cover one of these tools: dmalloc.

Dmalloc The debug memory allocation (dmalloc) library created by Gray Watson replaces malloc, realloc, calloc, free, and other memory management features with routines that provide debugging facilities that you can configure at runtime. The library has been tested on a variety of platforms. Follow the installation directions provided at https://dmalloc.com/ to configure, build, and install the library. Listing 6-11 shows you how to enable dmalloc to report the file and line numbers of calls that generate problems. This listing contains a short program that prints out some usage information and exits (it would typically be part of a longer program). Include the lines shown in bold font to allow dmalloc to report the file and line numbers of calls that cause problems. #include #include #include #ifdef DMALLOC #include "dmalloc.h" #endif void usage(char *msg) { fprintf(stderr, "%s", msg); free(msg); return; } int main(int argc, char *argv[]) { if (argc != 3 && argc != 4) { /* the error message won't be more than 80 chars */ char *errmsg = (char *)malloc(80); sprintf( errmsg, "Sorry %s,\nUsage: caesar secret_file keys_file [output_file]\n", getenv("USER") ); usage(errmsg); free(errmsg); exit(EXIT_FAILURE); }

116   Chapter 6

//---snip-- exit(EXIT_SUCCESS); } Listing 6-11: Catching a memory bug with dmalloc

I’ll show the output in a moment, but we need to discuss a few things first. The dmalloc distribution also comes with a command line utility. You can get further information on how to use this utility by running the following: % dmalloc --usage

Before debugging your program with dmalloc, enter the following at the command line: % dmalloc -l logfile -i 100 low

This command will set the logfile name to logfile, and instruct the library to perform a check every 100 invocations, specified in the -i argument. If you specify a larger number as the -i argument, dmalloc will check the heap less often, and your code will run faster; lower numbers are more likely to catch memory problems. The third command line argument enables a low number of debug features. Other options include runtime for minimal checking, or medium or high for more extensive heap verification. After executing this command, we can compile the program by using GCC as follows: % gcc -DDMALLOC caesar.c -ocaesar -ldmalloc

When you run the program, you should see the following error: % ./caesar Sorry student, Usage: caesar secret_file keys_file [output_file] debug-malloc library: dumping program, fatal error Error: tried to free previously freed pointer (err 61) Aborted (core dumped)

And if you examine the logfile, you’ll find the following information: % more logfile 1571549757: 3: Dmalloc version '5.5.2' from 'https://dmalloc.com/' 1571549757: 3: flags = 0x4e48503, logfile 'logfile' 1571549757: 3: interval = 100, addr = 0, seen # = 0, limit = 0 1571549757: 3: starting time = 1571549757 1571549757: 3: process pid = 29531 1571549757: 3: error details: finding address in heap 1571549757: 3: pointer '0x7ff010812f88' from 'caesar.c:29' prev access 'unknown' 1571549757: 3: ERROR: free: tried to free previously freed pointer (err 61) Dynamically Allocated Memory   117

These messages indicate that we’ve attempted to free the storage referenced by errmsg twice, first in the usage function and then in main, which constitutes a double-free vulnerability. Of course, this is just a single example of the types of bugs that dmalloc can detect, and other defects exist in the simple program we are testing. We’ll discuss other dynamic analysis tools in Chapter 11, along with further suggestions for how to use them.

Safety-Critical Systems Systems with high safety requirements frequently ban the use of dynamic memory, because memory managers can have unpredictable behavior that significantly impacts performance. Forcing all applications to live within a  fixed, pre-allocated area of memory can eliminate many of these problems and make it easier to verify memory use. In the absence of recursion, alloca, and VLAs (also prohibited in safety-critical systems), an upper bound on the use of stack memory can be derived statically, making it possible to prove that sufficient storage exists to execute the functionality of the application for all possible inputs. GCC also has a -Wvla flag that warns if a VLA is used, and a -Wvla-larger  -than=byte-size flag that warns for declarations of VLAs whose size is either unbounded, or bounded by an argument that allows the array size to exceed byte-size bytes.

Exercises Try these code exercises on your own: 1. Repair the use-after-free defect from Listing 6-4. 2. Perform additional testing of the program from Listing 6-11 by using dmalloc. Try varying inputs to the program to identify other memory management defects.

Summary In this chapter, you learned about working with memory that has allocated storage duration, and how this differs from objects of either automatic or static storage duration. We described the heap and memory managers and each of the standard memory management functions. We identified some common causes of errors when using dynamic memory, such as leaks, double-free vulnerabilities, and some mitigations to help avoid these problems. We also covered some more specialized memory allocation topics such as flexible array members, the alloca function, and variable-length arrays. We concluded the chapter with a discussion of debugging allocated storage problems by using the dmalloc library. In the next chapter, you’ll learn about characters and strings.

118   Chapter 6

7 CHAR ACTERS AND STRINGS

Strings are such an important and useful data type that nearly every programming language implements them in some form. Often used to represent text, strings constitute most of the data exchanged between an end user and a program, including text input fields, command line arguments, environment variables, and console input. In C, the string data type is modeled on the idea of a formal string (Hopcroft 1979): Let Σ be a non-empty finite set of characters, called the alphabet. A string over Σ is any finite sequence of characters from Σ. For example, if Σ = {0, 1}, then 01011 is a string over Σ.

In this chapter, we’ll talk about the various character sets, including ASCII and Unicode, that can be used to compose strings (the alphabet from

the formal definition). We’ll cover how strings are represented and manipulated using the legacy functions from the C Standard Library, the boundschecked interfaces, and POSIX and Windows APIs.

Characters The characters that people use to communicate aren’t naturally understood by digital systems, which operate on bits. To process characters, digital systems use character encodings that assign unique integer values, referred to as code points, to designate specific characters. As you’ll see, there are multiple ways to encode the same notional character in your program. Common standards used by C implementations for encoding characters include Unicode, ASCII, Extended ASCII, ISO 8859-1, Shift-JIS, and EBCDIC.9

ASCII The 7-bit American Standard Code for Information Interchange, better known as 7-bit ASCII, specifies a set of 128 characters and their coded representation (ANSI X3.4-1986). Characters up through 0x1f are control characters, such as null, backspace, and horizontal tab. Characters from 0x20 through 0x7e are all printable characters such as letters, digits, and symbols. We often refer to this standard with the updated name US-ASCII to clarify that this system was developed in the United States and focuses on the typographical symbols predominantly used in this country. Most modern character encoding schemes are based on US-ASCII, although they support many additional characters. Characters in the 0x80- 0xFF range are not defined by US-ASCII but are part of the 8-bit character encoding known as Extended ASCII. Numerous encodings for these ranges exist, and the actual mapping depends on the code page. A code page is a character encoding that maps a set of printable characters and control characters to unique numbers.

Unicode Unicode has become the universal character encoding standard for representing text in computer processing. It supports a much wider range of characters than ASCII does; the current Unicode Standard (Unicode 2020) encodes characters in the range U+0000 to U+10FFFF, which amounts to a 21-bit code space. An individual Unicode value is expressed as U+ followed by four or more hexadecimal digits in printed text. The Unicode characters U+0000 to U+007F are identical to those in US-ASCII, and the range U+0000 to U+00FF is identical to ISO 8859-1 (Latin-1), consisting of characters from the Latin script used throughout the Americas, Western Europe, Oceania, and much of Africa.

9. Unicode and ASCII are explicitly referenced by the C Standard.

120   Chapter 7

Unicode organizes code points into planes, which are continuous groups of 65,536 code points. There are 17 planes, identified by the numbers 0 to 16. The most commonly used characters, including those found in major, older encoding standards, have been placed into the first plane (0x0000 to 0xFFFF), which is called the basic multilingual plane (BMP), or Plane 0. Unicode also specifies several Unicode transformation formats (UTFs). UTFs are character encoding formats that assign each Unicode scalar value to a unique code unit sequence. A Unicode scalar value is any Unicode code point except high-surrogate and low-surrogate code points. A code unit is the minimal bit combination that can represent encoded text for processing or interchange. The Unicode Standard defines three UTFs to allow for code units of various sizes: UTF-8  Represents each character as a sequence of one to four 8-bit code units UTF-16  Represents each character as a sequence of one or two 16-bit code units UTF-32  Represents each character as a single 32-bit code unit The UTF-8 encoding is the dominant encoding for POSIX operating systems. It has the following desirable properties: •

• • •

It encodes US-ASCII characters (U+0000 to U+007F) as single bytes in the range 0x00 to 0x7F. This means that files and strings that contain only 7-bit ASCII characters have the same encoding under both ASCII and UTF-8. Using a null byte to terminate a string (a topic we’ll discuss later) works the same as for an ASCII string. All currently defined Unicode code points can be encoded using between 1 and 4 bytes. Unicode is designed to make it easy to identify character boundaries by scanning for well-defined bit patterns in either direction.

On Windows, you can compile and link your programs with the Visual C++ /utf8 flag to set the source and execution character sets as UTF-8. You’ll also need to configure Windows to use Unicode UTF-8 for worldwide language support, unless this changes in the future. UTF-16 is currently the dominant encoding for Windows operating systems. Like UTF-8, UTF-16 is a variable-width encoding. As just mentioned, the BMP consists of characters from U+0000 to U+FFFF. Characters whose code points are greater than U+FFFF are called supplementary characters. Supplementary characters are defined by a pair of code units called surrogates. The first code unit is from the high-surrogates range (U+D800- U+DBFF), and the second code unit is from the low-surrogates range (U+DC00- U+DFFF). UTF-32 is a fixed-length encoding, unlike other Unicode transformation formats, which are variable-length encodings. The main advantage of UTF-32 is that the Unicode code points are directly indexed, meaning that

Characters and Strings   121

you can find the nth code point in a sequence of code points in constant O(1) time. In contrast, a variable-length encoding requires accessing each code point sequentially to find the nth code point in a sequence.

Source and Execution Character Sets No universally accepted character encoding existed when C was originally standardized, so it was designed to work with a wide variety of character representations. Instead of specifying a character encoding like Java, each C implementation defines both a source character set in which source files are written and an execution character set used for character and string literals at compile time. Both the source and execution character sets must contain encodings for the uppercase and lowercase letters of the Latin alphabet; the 10 decimal digits; the 29 graphic characters; and the space, horizontal tab, vertical tab, form feed, and newline characters. The execution character set also includes alert, backspace, carriage return, and null characters. The character conversion and classification functions (such as isdigit) are evaluated at runtime, based on the locale-determined encoding in effect at the time of the call. Locales define local conventions of nationality, culture, and language.

Data Types C defines several data types to represent character data, some of which we have already seen. In particular, C offers the unadorned char type to represent narrow characters (those that can be represented in as few as 8 bits), and the wchar_t type to represent wide characters (those that may require more than 8 bits). char As we have already seen, char is an integer type, but each implementation defines whether it is signed or unsigned—meaning that in portable code, you cannot assume either. Use the char type for character data (where signedness has no meaning) and not for integer data (where signedness is important). The char type can be safely used to represent 7-bit character encodings, such as US-ASCII. For these encodings, the high-order bits are always 0, so you don’t have to be concerned about sign extension when a value of type char is converted to int and implementation defined as a signed type. The char type can also be used to represent 8-bit character encodings such as Extended ASCII, ISO/IEC 8859, EBCDIC, and UTF-8. These 8-bit character encodings can be problematic on implementations that define char as an 8-bit signed type. For example, the following code prints the string end of file when an EOF is detected: char c = 'ÿ'; // extended character if (c == EOF) puts("end of file");

122   Chapter 7

Assuming the implementation-defined execution character set is ISO/ IEC 8859-1, the Latin small letter y with diaeresis (ÿ) is defined to have the representation 255 (0xFF). For implementations in which char is defined as a signed type, c will be sign-extended to the width of signed int, making the ÿ character indistinguishable from EOF because they now have the same representation. A similar problem occurs when using the character classification functions defined in . These library functions accept a character argument as an int or the value of the macro EOF, and return true if the value belongs to a set of characters defined by the description of the function. For example, the isdigit function tests whether the character is a decimal-digit character in the current locale. Any argument value that isn’t a valid character or EOF will result in undefined behavior. To avoid undefined behavior when invoking these functions, cast c to unsigned char before the integer promotions, as shown here: char c = 'ÿ'; if (isdigit((unsigned char)c)) { puts("c is a digit"); }

The value stored in c is zero-extended to the width of signed int, eliminating the undefined behavior because the resulting value can still be represented as an unsigned char. int Use the int type for data that could be either EOF (a negative value) or character data interpreted as unsigned char and then converted to int. This type is returned by functions, such as fgetc, getc, getchar, and ungetc, that read character data from a stream. As we’ve seen, this type is also accepted by the character-handling functions from , because they might well be passed the result of fgetc or related functions. wchar_t The wchar_t type is an integer type added to C to process the characters of a large character set. It can be a signed or unsigned integer type, depending on the implementation, and has an implementation-defined inclusive range of WCHAR_MIN to WCHAR_MAX. Most implementations define wchar_t to be either a 16- or 32-bit unsigned integer type, but implementations that don’t support localization may define wchar_t to have the same width as char. C does not permit a variable-length encoding for wide strings (despite UTF-16 being used in practice on Windows). Implementations can conditionally define the macro __STDC_ISO_10646__ as an integer constant of the form yyyymmL (for example, 199712L) to mean that the wchar_t type is used to represent Unicode characters corresponding to the specified version of the standard. Implementations that chose a 16-bit type for wchar_t cannot meet the requirements for defining __STDC_ISO_10646__ for ISO/IEC 10646 editions Characters and Strings   123

more recent than Unicode 3.1 (ISO/IEC 10646-1:2000 and ISO/IEC 10646-2:2001). Consequently, the requirement for defining __STDC_ISO  _10646__ is either a wchar_t type larger than 20 bits, or a 16-bit wchar_t and a value for __STDC_ISO_10646__ earlier than 200103L. The wchar_t type can be used for encodings other than Unicode, such as wide EBCDIC. Writing portable code using wchar_t can be difficult because of the range of implementation-defined behavior. For example, Windows uses a 16-bit unsigned integer type, while Linux typically uses a 32-bit unsigned integer type. Code that calculates the lengths and sizes of wide-character strings is error-prone and must be performed with care. char16_t and char32_t Newer languages (including Ada95, Java, TCL, Perl, Python, and C#) have data types for Unicode characters. C11 introduced the 16- and 32-bit character data types char16_t and char32_t, declared in , to provide data types for UTF-16 and UTF-32 encodings, respectively. C11 doesn’t include library functions for the new data types, except for one set of character conversion functions, to allow library developers to implement them. Without library functions, these types have limited usefulness. C defines two environment macros that indicate how characters represented in these types are encoded. If the environment macro __STDC_UTF_16__ has the value 1, values of type char16_t are UTF-16 encoded. If the environment macro __STDC_UTF_32__ has the value 1, values of type char32_t are UTF32 encoded. If the macro isn’t defined, another implementation-defined encoding is used. Visual C++ does not define these macros.

Character Constants C allows you to specify character constants, also known as character literals, which are sequences of one or more characters enclosed in single quotes, such as 'ÿ'. Character constants allow you to specify character values in the source code of your program. Table 7-1 shows the types of character constants that can be specified in C. Table 7-1: Types of Character Constants

Prefix

Type

None

int

L'a'

The unsigned type corresponding to wchar_t

u'a'

char16_t

U'a'

char32_t

The oddest thing in Table 7-1 is that a nonprefixed character constant, such as 'a', has type int rather than type char. In C, for historical reasons, if a character constant contains only a single character or escape sequence, the value of the character constant is represented as an object with type

124   Chapter 7

char converted to type int. This differs from C++, in which a character literal that contains only one character has type char.

The value of a character constant containing more than one character (for example, 'ab') is implementation defined. So is the value of a source character that cannot be represented as a single code unit in the execution character set. The earlier example of 'ÿ' is one such case. If the exe­cution character set is UTF-8, the value might be 0xC3BF to reflect the UTF-8 encoding of the two code units needed to represent the U+00FF code-point value. C2x will add the u8 prefix for character literals to represent a UTF-8 encoding. Until the release of C2x, there will not be a character-literal prefix for UTF-8 characters, unless implementations decide to implement it in advance.

Escape Sequences The single quote (') and backslash (\) have special meanings, so they cannot be directly represented as characters. Instead, to represent the single quote, we use the escape sequence \', and to represent the backslash, we use \\. We can represent other characters, such as the question mark (?), and arbitrary integer values by using the escape sequences shown in Table 7-2. Table 7-2: Escape Sequences

Character

Escape sequence

Single quote

\'

Double quote

\"

Question mark

\?

Backslash

\\

Alert

\a

Backspace

\b

Form feed

\f

Newline

\n

Carriage return

\r

Horizontal tab

\t

Vertical tab

\v

Octal character

\up to 3 octal digits

Hexadecimal character

\x hexadecimal digits

The following nongraphical characters are represented by escape sequences consisting of the backslash followed by a lowercase letter: \a (alert), \b (backspace), \f (form feed), \n (newline), \r (carriage return), \t (horizontal tab), and \v (vertical tab). Octal digits can be incorporated into an octal escape sequence to construct a single character for a character constant, or a single wide character for a wide-character constant. The numerical value of the octal integer specifies the value of the desired character or wide character. A backslash followed by numbers is always interpreted as an octal value. For example, you can Characters and Strings   125

represent the backspace character (8 decimal) as the octal value \10 or, equivalently, \010. We can also incorporate the hexadecimal digits that follow the \x to construct a single character or wide character for a character constant. The numerical value of the hexadecimal integer forms the value of the desired character or wide character. For example, you can represent the backspace character as the hexadecimal value \x8 or, equivalently, \x08.

Linux Character encodings have evolved differently on various operating systems. Before UTF-8 emerged, Linux typically relied on various language-specific extensions of ASCII. The most popular of these were ISO 8859-1 and ISO 8859-2 in Europe, ISO 8859-7 in Greece, KOI-8/ISO 8859-5/CP1251 in Russia, EUC and Shift-JIS in Japan, and BIG5 in Taiwan. Linux distributors and application developers are phasing out these older legacy encodings in favor of UTF-8 to represent localized text strings (Kuhn 1999). GCC has several flags that allow you to configure character sets. Here are a couple of flags you may find useful: -fexec-charset=charset

The –fexec-charset flag sets the execution character set that’s used to interpret string and character constants. The default is UTF-8. The charset can be any encoding supported by the system’s iconv library routine described later in this chapter. For example, setting –fexec-charset=IBM1047 instructs GCC to interpret string constants hardcoded in source code, such as printf format strings, according to EBCDIC code page 1047. To select the wide execution character set, used for wide string and character constants, use the –fwide-exec-charset flag: -fwide-exec-charset=charset

The default is UTF-32 or UTF-16, corresponding to the width of wchar_t.

To set the input character set, used for translation from the character set of the input file to the source character set used by GCC, use the –finput-charset flag: -finput-charset=charset

Clang has –fexec-charset and –finput-charset, but not –fwide-exec-charset. Clang allows you to set charset to UTF-8 only and rejects any attempt to set it to something else.

Windows Support for character encodings in Windows has irregularly evolved. Programs developed for Windows can handle character encodings 126   Chapter 7

using either Unicode interfaces or interfaces that implicitly rely on locale-dependent character encodings. For most modern applications, you should choose the Unicode interfaces by default to ensure that the application behaves as you expect when processing text. Generally, this code will have better performance as narrow strings passed to Windows library functions are frequently converted to Unicode strings. The main and wmain Entry Points Visual C++ supports two entry points to your program: main, which allows you to pass narrow-character arguments, and wmain, which allows you to pass wide-character arguments. You declare formal parameters to wmain by using a format similar to main, as shown in Table 7-3. Table 7-3: Windows Program Entry Point Declarations

Narrow-character arguments

Wide-character arguments

int main(void); int main(int argc, char *argv[]); int main(int argc, char *argv[],  char *envp[]);

int wmain(void); int wmain(int argc, wchar_t *argv[]); int wmain(int argc, wchar_t *argv[],  wchar_t *envp[]);

For either entry point, the character encoding ultimately depends on the calling process. However, by convention, the main function generally receives its optional arguments and environment as pointers to text encoded with the current Windows (also called ANSI) code page, while the wmain function generally receives UTF-16 encoded text. When you run a program from a shell such as the command prompt, the shell’s command interpreter converts the arguments into the proper encoding for that entry point. A Windows process starts with a UTF-16 encoded command line. The startup code emitted by the compiler/linker calls the CommandLineToArgvW function to convert the command line to the argv form required to call main, or passes the command line arguments directly to the argv form required to call wmain. In a call to main, the results are then transcoded to the current Windows code page, which can vary from system to system. The ASCII character ? is substituted for characters that lack representation in the current Windows code page. The Windows console uses an original equipment manufacturer (OEM) code page when writing data to the console. The actual encoding used varies from system to system, but is often different from the Windows code page. For example, on a US English version of Windows, the Windows code page may be Windows Latin 1, while the OEM code page may be DOS Latin US. In general, writing textual data to stdout or stderr requires the text to be converted to the OEM code page first, or requires setting the console’s output code page to match the encoding of the text being written out. Failure to do so may cause unexpected output to be printed to the console. However, even if you carefully match the character encodings between your program and the console, the console might still fail to display the characters as expected because of other factors, such as the current font selected Characters and Strings   127

for the console not having the appropriate glyphs required to represent the characters. Additionally, the Windows console has historically been unable to display characters outside of the Unicode BMP because it stores only a 16-bit value for character data for each cell. Narrow versus Wide Characters There are two versions of all system APIs in the Win32 SDK: a narrow Windows (ANSI) version with an A suffix, and a wide character version with a W suffix: int SomeFuncA(LPSTR SomeString); int SomeFuncW(LPWSTR SomeString);

You should determine whether your app is going to use wide (UTF-16) or narrow characters and then code accordingly. The best practice is to explicitly call the narrow or wide string version of each function and pass a string of the appropriate type: SomeFuncW(L"String"); SomeFuncA("String");

Examples of actual functions from the Win32 SDK include the MessageBoxA/MessageBoxW and CreateWindowExA/CreateWindowExW functions.

Character Conversion Although international text is increasingly encoded in Unicode, it is still encoded in language- or country-dependent character encodings, making it necessary to convert between these encodings. In particular, Windows still operates in locales with traditional, limited character encodings, such as IBM EBCDIC and ISO 8859-1. Programs frequently need to convert between the Unicode and traditional encoding schemes when performing I/O. It is not possible to convert all strings to each language- or countrydependent character encoding. This is obvious when the encoding is US-ASCII, which can’t represent any character requiring more than seven bits of storage. Latin-1 will never encode the character 晥 properly, and many kinds of non-Japanese letters and words cannot be converted to Shift-JIS without losing information. The following sections describe various approaches to converting between character encodings. C Standard Library The C Standard Library provides a handful of functions to convert between narrow code units (char) and wide code units (wchar_t). The mbtowc (multibyte to wide character), wctomb (wide character to multibyte), mbrtowc (multibyte restartable to wide character), and wcrtomb (wide character restartable to multibyte) functions convert one code unit at a time, ­writing the result to an output object or buffer. The mbstowcs (multibyte 128   Chapter 7

string to wide character string), wcstombs (wide character string to multibyte string), mbsrtowcs (multibyte string restartable to multibyte string), and wcsrtombs (wide character string restartable to multibyte string) functions convert strings of code units at a time, writing the result to an output buffer. Conversion functions need to store data to properly process a sequence of conversions between function calls. The nonrestartable forms store the state internally. The restartable versions have an additional parameter that is a pointer to an object of type mbstate_t that describes the current conversion state of the associated multibyte character sequence. This object holds the state data that makes it possible to restart the conversion where it left off after another call to the function to perform an unrelated conversion. The string versions are for performing bulk conversions of multiple code units at once. These functions have a few limitations. As discussed earlier, Windows uses 16-bit code units for wchar_t. This can be a problem, because the C Standard requires an object of type wchar_t to be capable of representing any character in the current locale, and a 16-bit code unit can be too small to do so. C technically doesn’t allow you to use multiple objects of type wchar_t to represent a single character. Consequently, the standard conversion functions may result in a loss of data. On the other hand, most POSIX implementations use 32-bit code units for wchar_t, allowing the use of UTF-32. Because a single UTF-32 code unit can represent a whole code point, conversions using standard functions cannot lose or truncate data. The C Standards committee added the following functions to C11 to address the potential loss of data using standard conversion functions: mbrtoc16, c16rtomb  Converts between a sequence of narrow code units and one or more char16_t code units mbrtoc32, c32rtomb  Converts a sequence of narrow code units to one or more char32_t code units

The first two functions convert between locale-dependent character encodings, represented as an array of char, and UTF-16 data stored in an array of char16_t (assuming __STDC_UTF_16__ has the value 1). The second two functions convert between the locale-dependent encodings and UTF-32 data stored in array of char32_t encoded data (assuming __STDC_UTF_32__ has the value 1). The program shown in Listing 7-1 uses the mbrtoc16 function to convert a UTF-8 input string to a UTF-16 encoded string. #include #include #include #include #include





u #if __STDC_UTF_16__ != 1 #error "__STDC_UTF_16__ not defined"

Characters and Strings   129

#endif int main(void) { v setlocale(LC_ALL, "en_US.utf8"); char input[] = u8"I ♥️ s!"; const size_t input_size = sizeof(input); char16_t output[input_size]; // UTF-16 requires less code units than UTF-8 char *p_input = input; char *p_end = input + input_size; char16_t *p_output = output; size_t code; mbstate_t state = {0}; puts(input); w while ((code = mbrtoc16(p_output, p_input, p_end-p_input, &state))){ if (code == (size_t)-1) break; // -1 – invalid code unit sequence detected else if (code == (size_t)-2) break; // -2 – code unit sequence missing elements else if (code == (size_t)-3) p_output++; // -3 – high surrogate from a surrogate pair else { p_output++; // one value was written out p_input += code; // code is the # of code units read by function } } size_t output_size = p_output - output + 1; printf("Converted to %zu UTF-16 code units: [ ", output_size); for(size_t x = 0; x < output_size; ++x) printf("%#x ", output[x]); puts("]"); } Listing 7-1: Using the mbrtoc16 function to convert a UTF-8 string to a char16_t string

We call the setlocale function v to set the multibyte character encoding to UTF-8 by passing an implementation-defined string. The preprocessing directives u ensure that the macro __STDC_UTF_16__ has the value 1. (Refer to Chapter 9 for more information on preprocessing directives.) As a result, each call to the mbrtoc16 function converts a single code point from a UTF-8 representation to a UTF-16 representation. If the resulting UTF-16 code unit is a high surrogate (from a surrogate pair), the state object is updated to indicate that the next call to mbrtoc16 will write out the low surrogate without considering the input string. There is no string version of the mbrtoc16 function, so we loop through a UTF-8 input string iteratively, calling the mbrtoc16 function w, to convert it to a UTF-16 string. In the case of an encoding error, the mbrtoc16 function returns (size_t)-1 and, if the code unit sequence is missing elements, it returns (size_t)-2. If either situation occurs, the loop terminates and the conversion ends. A return value of (size_t)-3 means that the function output the high surrogate from a surrogate pair, and then stored an indicator in the state parameter. The indicator is used the next time the mbrtoc16 function is called so it can output the low surrogate from a surrogate pair to form a 130   Chapter 7

complete char16_t sequence that represents a single code point. All restartable encoding conversion functions in the C Standard behave similarly with the state parameter. If the function returns anything other than (size_t)-1, (size_t)-2, or (size_t)-3, the p_output pointer is incremented and the p_input pointer is increased by the number of code units read by the function, and the conversion of the string continues. libiconv GNU libiconv is a commonly used cross-platform, open source library for performing string encoding conversions. It includes the iconv_open function that allocates a conversion descriptor you can use to convert byte sequences from one character encoding to another. The documentation for this function, found at https://www.gnu.org/, defines strings you can use to identify a particular charset such as ASCII, ISO−8859−1, SHIFT_JIS, or UTF−8 to denote the locale-dependent character encoding. Win32 Conversion APIs The Win32 SDK provides two functions for converting between wide and narrow character strings: MultiByteToWideChar  Maps a character string to a new UTF-16 (wide

character) string WideCharToMultiByte  Maps a UTF-16 (wide character) string to a new

character string The MultiByteToWideChar function maps string data that’s encoded in an arbitrary character code page to a UTF-16 string. Similarly, the function WideCharToMultiByte maps string data encoded in UTF-16 to an arbitrary character code page. Because UTF-16 data cannot be represented by all code pages, this function can specify a default character to use in place of any UTF-16 character that cannot be converted.

Strings C doesn’t support a primitive string type and probably never will. Instead, it implements strings as arrays of characters. C has two types of strings: narrow and wide. A narrow string has the type array of char. It consists of a contiguous sequence of characters that include a terminating null character. A pointer to a string points to its initial character. The size of a string is the number of bytes allocated to the backing array storage. The length of a string is the number of code units (bytes) preceding the first null character. In Figure 7-1, the size of the string is 7, and the length of the string is 5. Elements of the backing array beyond the last element must not be accessed. Elements of the array that haven’t been initialized must not be read. Characters and Strings   131

Size h

e

l

l

o \0 ??

Length

Figure 7-1: Sample narrow string

A wide string has the type array of wchar_t. It’s a contiguous sequence of wide characters that include a terminating null wide character. A pointer to a wide string points to its initial wide character. The length of a wide string is the number of code units preceding the first null wide character. Figure 7-2 illustrates both the UTF-16BE (big endian) and UTF-16LE (little endian) representation of hello. The size of the array is implementationdefined. This array is 14 bytes and assumes an implementation that has an 8-bit byte and 16-bit wchar_t type. The length of this string is 5, as the number of characters has not changed. Elements of the backing array beyond the last element must not be accessed. Elements of the array that haven’t been initialized must not be read. The size of a wide string is the number of bytes allocated to the backing array storage. A pointer to a wide string points to its initial (lowest addressed) wide character.

UTF-16BE

\0

h

\0

e

\0

1

\0

1

\0

o

\0 \0 ?? ??

h

\0

e

\0

1

\0

1

\0

o

\0

\0 \0 ?? ??

UTF-16LE The length of a wide string is the number of wide characters preceding the null wide character.

Figure 7-2: Sample UTF-16LE wide string

String Literals A character string literal is a string constant represented by a sequence of zero or more multibyte characters enclosed in double quotes; for example, "ABC". You can use various prefixes to declare string literals of different character types: • • • • •

char string literal type, such as "ABC" wchar_t string literal type with L prefix, such as L"ABC"

UTF-8 string literal type with u8 prefix, such as u8"ABC" char16_t string literal type with u prefix, such as u"ABC" char32_t string literal type with U prefix, such as U"ABC"

The C Standard doesn’t mandate that an implementation use ASCII for string literals. However, you can use the u8 prefix to force a string literal to be UTF-8 encoded, and if all the characters in the literal are ASCII

132   Chapter 7

characters, the compiler will produce an ASCII string literal, even if the implementation would normally encode string literals in another encoding (for example, EBCDIC). A string literal has a non- const array type. Modifying a string literal is undefined behavior and prohibited by the CERT C rule STR30-C (Do not attempt to modify string literals). This is because these string literals may be stored in read-only memory, or multiple string literals may share the same memory, resulting in multiple strings being altered if one string is modified. String literals often initialize array variables, which you can declare with an explicit bound that matches the number of characters in the string literal. Consider the following declaration: #define S_INIT "abc" // ---snip--const char s[4] = S_INIT;

The size of the array s is four, the exact size required to initialize the array to the string literal, including the space for a trailing null byte. If you add another character to the string literal used to initialize the array, however, the meaning of the code changes substantially: #define S_INIT "abcd" // ---snip--const char s[4] = S_INIT;

The size of the array s remains four, although the size of the string literal is now five. As a result, the array s is initialized to the character array "abcd" with the trailing null byte omitted. This syntax is by design, to allow you to initialize a character array and not a string. Therefore, it is unlikely that your compiler will diagnose this declaration as an error. GCC issues a warning when such an array is passed as an argument to a string function. There is some risk that, if the string literal changes during maintenance, a string could unintentionally be changed to a character array with no terminating null character, particularly when the string literal is defined separately from the declaration, as in this example. If your intent is to always initialize s to a string, you should omit the array bound. If you don’t specify the bound of the array, the compiler will allocate sufficient space for the entire string literal, including the terminating null character: const char s[] = S_INIT;

This approach simplifies maintenance because the size of the array can always be determined even if the size of the string literal changes. The size of arrays declared using this syntax can be determined at compile time by using the sizeof operator: size_t size = sizeof s;

Characters and Strings   133

If, instead, we declared this string as follows const char *foo = S_INIT;

we would need to invoke the strlen function to get the following length size_t size = strlen(foo) + 1U;

which may incur a runtime cost.

String-Handling Functions Several approaches can be used to manage strings in C, the first of which are the C Standard Library functions. Narrow string-handling functions are defined in the header file, and wide string-handling functions in . These legacy string-handling functions have been associated in recent years with various security vulnerabilities. This is because they don’t check the size of the array (frequently lacking the information needed to perform such checks) and trust you to provide adequately sized character arrays to hold the output. While it is possible to write safe, robust, and error-free code using these functions, the library promotes programming styles that can result in buffer overflows if a result is too large for the provided array. These functions aren’t inherently insecure but are prone to misuse and need to be used carefully (or not at all). As a result, C11 introduced the normative (but optional) Annex K bounds-checking interfaces. This annex provides alternative library functions intended to promote safer, more secure programming by requiring you to provide the length of output buffers, for example, and validating that these buffers are adequately sized to contain the output from these functions. For instance, Annex K defines the strcpy_s, strcat_s, strncpy_s, and strncat_s functions as close replacements for the C Standard Library’s strcpy, strcat, strncpy, and strncat functions. POSIX also defines several string-handling functions, such as strdup and strndup, that provide another set of string APIs you can use on POSIXcompliant platforms such as Linux and Unix (IEEE Std 1003.1:2018). Visual C++ provides all the string-handling functions defined by the C Standard Library up to C99 but doesn’t implement the full POSIX specification.10 Visual C++ names many of its runtime library functions with a leading underscore; for example, _strdup instead of strdup. Visual C++ also supports many of the safe string-handling functions from Annex K and will diagnose use of an unsafe variant, unless you define _CRT_SECURE_NO_WARNINGS prior to including the header file that declares the function. We’ll discuss each of these string-handling libraries in the subsequent sections.

10. For more information on Microsoft’s POSIX support, see Compatibility at https://docs.microsoft.com/en-us/cpp/c-runtime-library/compatibility?view=vs-2019/.

134   Chapter 7

and The C Standard Library includes well-known functions such as strcpy, strncpy, strcat, strncat, strlen, and so forth, as well as the memcpy and ­memmove functions you can use to copy and move strings, respectively. The C Standard also provides a wide character interface that operates on objects of type wchar_t instead of char. (These function names are similar to the narrow string function names, except that str is replaced with wc, and a w is added in front of the memory function names.) Table 7-4 gives some examples of narrow and wide character string functions. It’s important not to confuse them. Table 7-4: Narrow and Wide String Functions

Narrow (char)

Wide ( wchar_t)

Description

strcpy

wcscpy

String copy

strncpy

wcsncpy

Truncated, zero-filled copy

memcpy

wmemcpy

Copies a specified number of code units

memmove

wmemmove

Copies a specified number of (possibly overlapping) code units

strcat

wcscat

Concatenates strings

strncat

wcsncat

Concatenates strings with truncation

strcmp

wcscmp

Compares strings

strncmp

wcsncmp

Compares strings to a point

strchr

wcschr

Locates a character in a string

strcspn

wcscspn

Computes the length of a complementary string segment

strpbrk

wcspbrk

Finds the first occurrence of a set of characters in a string

strrchr

wcsrchr

Finds the first occurrence of a character in a string

strspn

wcsspn

Computes the length of a string segment

strstr

wcsstr

Finds a substring

strtok

wcstok

String tokenizer (modifies the string being tokenized)

memchr

wmemchr

Finds a code unit in memory

strlen

wcslen

Computes string length

memset

wmemset

Fills memory with a specified code unit

These string-handling functions are considered efficient because they leave memory management to the caller and can be used with both statically and dynamically allocated storage. In the next couple of sections, I’ll go into more detail on some of the most commonly used of these functions.

Characters and Strings   135

Size and Length As mentioned earlier in this chapter, strings have both a size (which is the number of bytes allocated to the backing array storage) and a length. You can determine the size of a statically allocated backing array at compile time by using the sizeof operator: char str[100] = "Here comes the sun"; size_t str_size = sizeof(str); // str_size is 100

You can compute the length of a string by using the strlen function: char str[100] = "Here comes the sun"; size_t str_len = strlen(str); // str_len is 18

The wcslen function computes the length of a wide string measured by the number of code units preceding the terminating null wide character: wchar_t str[100] = L"Here comes the sun"; size_t str_len = wcslen(str); // str_len is 18

The length is a count of something, but what exactly is being counted can be unclear. Here are some of the things that could be counted when taking the length of a string: Bytes  Useful when allocating storage. Code units  Number of individual code units used to represent the string. This length depends on encoding and can also be used to allocate memory. Code points  Code points (characters) can take up multiple code units. This value is not useful when allocating storage. Extended grapheme cluster  A group of one or more Unicode scalar values that approximates a single user-perceived character. Many individual characters, such as “é”, “김”, and “🇮🇳” may be constructed from multiple Unicode scalar values. Unicode’s boundary algorithms combine these code points into extended grapheme clusters. The strlen and wcslen functions count code units. For the strlen function, this corresponds to the number of bytes. Determining the amount of storage required by using the wcslen function is more complicated because the size of the wchar_t type is implementation defined. Listing 7-2 contains examples of dynamically allocating storage for both narrow and wide strings. // narrow strings char str1[] = "Here comes the sun"; char *str2 = malloc(strlen(str1) + 1)); 

136   Chapter 7

// wide strings wchar_t wstr1[] = L"Here comes the sun"; wchar_t *wstr2 = malloc((wcslen(wstr1) + 1) * sizeof(wchar_t)); Listing 7-2: Dynamically allocating storage for narrow and wide string functions

For narrow strings, we can determine the size of the string by using the strlen function and then adding 1 for the terminating null character before allocating. For wide strings, we can determine the size of the string by using the wcslen function, add 1 to accommodate the terminating null wide character, and then multiply the result of this operation by the size of the wchar_t type. Code point or extended grapheme clusters counts cannot be used for storage allocation, because they consist of an unpredictable number of code units.11 Extended grapheme clusters are used to determine where to truncate a string, for example, because of a lack of storage. Truncation at extended grapheme cluster boundaries avoids slicing user-perceived characters. Calling the strlen function can be an expensive operation because it needs to traverse the length of the array looking for a null character. The following is a straightforward implementation of the strlen function: size_t strlen(const char * str) { const char *s; for (s = str; *s; ++s) {} return s - str; }

The strlen function has no way of knowing the size of the object referenced by str. If you call strlen with an invalid string that lacks a null character before the bound, the function will access the array beyond its end, resulting in undefined behavior. Passing a null pointer to strlen will also result in undefined behavior (a null-pointer dereference). This implementation of the strlen function also has undefined behavior for strings larger than PTRDIFF_MAX. You should refrain from creating such objects (in which case this implementation is fine). The strcpy Function Calculating the size of dynamically allocated memory is not as easy. One approach is to store the size when allocating and reuse this value later. The code snippet in Listing 7-3 uses the strcpy function to make a copy of str by determining the length and then adding 1 to accommodate the terminating null character: char str[100] = "Here comes the sun"; size_t str_size = strlen(str) + 1;

11.For an interesting exposition on string length, see “It’s Not Wrong that “ ”.length == 7” at https://hsivonen.fi/string-length/. Characters and Strings   137

char *dest = (char *)malloc(str_size); if (dest) { strcpy(dest, str); } else { /* handle error */ } Listing 7-3: Copying a string

We can then use str_size to dynamically allocate the storage for the copy. The strcpy function copies the string from the source string (str) to the destination string (dest), including the terminating null character. The strcpy function returns the address of the beginning of the destination string, which is ignored in this example. A common implementation of the strcpy function is as follows: char *strcpy(char *dest, const char *src) { char *save = dest; while (*dest++ = *src++); return save; }

This code saves a pointer to the destination string in save (to use as the return value) before copying all the bytes from the source to the destination array. The while loop terminates when the first null byte is copied. Because strcpy doesn’t know the length of the source string or the size of the destination array, it assumes that all the function’s arguments have been validated by the caller, allowing the implementation to simply copy each byte from the source string to the destination array without checks. Argument Checking Argument checking can be performed by either the calling function or the called function. Redundant argument testing by both the caller and the callee is a largely discredited style of defensive programming. The usual discipline is to require validation on only one side of each interface. The most time-efficient approach is for the caller to perform the check, because the caller should have a better understanding of the program state. In Listing 7-3, we can see that the arguments to strcpy are valid without introducing further redundant tests: the variable str references a statically allocated array that was properly initialized in the declaration, and the dest parameter is a non-null pointer referencing dynamically allocated storage of sufficient size to hold a copy of str, including the null character. Therefore, the call to strcpy is safe, and the copy can be performed in a time-efficient manner. This approach to argument checking is commonly used by C Standard Library functions because it adheres to the “spirit of C,” in that it’s optimally efficient and trusts the programmer to pass valid arguments.

138   Chapter 7

The safer, more secure, space-efficient approach is for the callee to check the arguments. This approach is less error-prone because the library function implementer validates the arguments, so we no longer need to trust the programmer to pass valid ones. The function implementer is usually in a better position to understand which arguments need to be validated. If the input validation code is defective, the repair needs to be made in only one place. All the code to validate the arguments is in one place, so this approach is typically more space-efficient. However, because these tests run even when unnecessary, they can also be less time-efficient. Frequently, the caller of these functions will place checks before suspect system calls that may or may not already perform similar checks. This approach would also impose additional error handling on callees that don’t currently return error indications but would presumably need to if they validated arguments. For strings, the called function can’t always determine whether the argument is a valid nullterminated string, or points to sufficient space to make a copy. The lesson here is don’t assume that the C Standard Library functions validate arguments unless the standard explicitly requires them to. The memcpy Function The memcpy function copies a specified number of characters, size, from the object referenced by src into the object referenced by dest: void *memcpy(void * restrict dest, const void * restrict src, size_t size);

You can use the memcpy function instead of strcpy to copy strings when the size of the destination array is larger than or equal to the size argument to memcpy, the source array contains a null character before the bound, and the string length is less than size - 1 (so that the resulting string will be properly null-terminated). The best advice is to use strcpy when copying a string, and memcpy when copying only raw, untyped memory. Also remember that the assignment (=) operator can efficiently copy objects in many cases. Most of the C Standard Library functions return a pointer to the beginning of the string passed as an argument so that you can nest calls to string functions. For example, the following sequence of nested function calls constructs a person’s full legal name by copying, then concatenating, the constituent parts: strcat(strcat(strcat(strcat(strcpy(full, first), " "), middle), " "), last);

However, piecing together the array full from its substrings requires this string to be scanned many more times than necessary; it would have been more useful for the functions to return pointers to the end of the modified string, to eliminate this need for rescanning. C2x will introduce a string-copy function with a better interface design, memccpy. POSIX environments should already provide this, but you may need to enable its declaration as follows: #define _XOPEN_SOURCE 700 #include Characters and Strings   139

The gets Function The gets function is a flawed input function that accepts input without providing any way to specify the size of the destination array. For that reason, it cannot prevent buffer overflows. As a result, the gets function was deprecated in C99 and eliminated from C11. However, it has been around for many years, and most libraries still provide an implementation for backward compatibility, so you may see it in the wild. You should never use this function, and you should replace any use of the gets function you find in any code you are maintaining. Because the gets function is so bad, we’ll spend some time examining why it’s so awful. The function shown in Listing 7-4 prompts the user to enter either y or n to indicate whether they’d like to continue. This function has undefined behavior if more than eight characters are entered at the prompt. However, the gets function has no way of knowing how large the destination array is and will simply write beyond the end of the array object. #include #include void get_y_or_n(void) { char response[8]; puts("Continue? [y] n: "); gets(response); if (response[0] == 'n') exit(0); return; } Listing 7-4: Misuse of the obsolete gets function

Listing 7-5 shows a simplified implementation of the gets function. As you can see, the caller of this function has no way to limit the number of characters read. char *gets(char *dest) { int c; char *p = dest; while ((c = getchar()) != EOF && c != '\n') { *p++ = c; } *p = '\0'; return dest; } Listing 7-5: gets function implementation

The gets function iterates reading a character at a time. The loop terminates if either an EOF or newline '\n' character is read. Otherwise, the function will continue to write to the dest array without concern for the boundaries of the object.

140   Chapter 7

Listing 7-6 shows the get_y_or_n function from Listing 7-4 with the gets function inlined. #include #include void get_y_or_n(void) { char response[8]; puts("Continue? [y] n: "); int c; char *p = response; u while ((c = getchar()) != EOF && c != '\n') { *p++ = c; } *p = '\0'; if (response[0] == 'n') exit(0); } Listing 7-6: Poorly written while loop

The size of the destination array is now available, but the while loop u doesn’t use this information. You should ensure that reaching the size of the destination array is a loop termination condition when reading or writing to an array in a loop such as this one.

Annex K Bounds-Checking Interfaces C11 introduced the Annex K bounds-checking interfaces with alternative functions that verify that output buffers are large enough for the intended result and return a failure indicator if they aren’t. These functions are designed to prevent writing data past the end of an array, and to nullterminate all string results. These string-handling functions leave memory management to the caller, and memory can be statically or dynamically allocated before the functions are invoked. Microsoft created the C11 Annex K functions to help retrofit its legacy code base in response to numerous, well-publicized security incidents in the 1990s. These functions were then proposed to the C Standards committee for standardization, published as ISO/IEC TR 24731-1 (ISO/IEC TR 24731-1:2007), and then later incorporated into C11 as a set of optional extensions. Despite the improved usability and security provided by these functions, they aren’t yet widely implemented at the time of writing. The gets_s Function The Annex K bounds-checking interface has a gets_s function we can use to eliminate the undefined behavior caused by the gets function in Listing 7-4, as shown in Listing 7-7. The two functions are similar, except that the gets_s function checks the array bounds. The default behavior that occurs when the maximum number of characters input is exceeded is implementation

Characters and Strings   141

defined, but typically the abort function is called. You can change this behavior via the set_constraint_handler_s function, which I’ll explain further in “Runtime Constraints” on page 143. #define __STDC_WANT_LIB_EXT1__ 1 #include #include void get_y_or_n(void) { char response[8]; size_t len = sizeof(response); puts("Continue? [y] n: "); gets_s(response, len); if (response[0] == 'n') exit(0); } Listing 7-7: Use of the gets_s function.

The first line in Listing 7-7 defines the __STDC_WANT_LIB_EXT1__ macro to expand to the value 1. We then include the header files that define the bounds-checking interfaces, allowing them to be used in your program. Unlike the gets function, the gets_s function takes a size argument. Consequently, the revised function calculates the size of the destination array by using the sizeof operator and passes this value as an argument to the gets_s function. The implementation-defined behavior is the result of the runtime-constraint violation. The strcpy_s Function The strcpy_s function is a close replacement for the strcpy function defined in . The strcpy_s function copies characters from a source string to a destination character array up to and including the terminating null character. Here’s the strcpy_s signature: errno_t strcpy_s( char * restrict s1, rsize_t s1max, const char * restrict s2 );

The strcpy_s function has an extra argument of type rsize_t that specifies the maximum length of the destination buffer. The strcpy_s function succeeds only when it can fully copy the source string to the destination without overflowing the destination buffer. The strcpy_s function verifies that the following runtime constraints aren’t violated: • • • • •

142   Chapter 7

Neither s1 nor s2 are null pointers. s1max is not > RSIZE_MAX. s1max does not equal zero. s1max is > strnlen_s(s2, s1max). Copying does not take place between objects that overlap.

To perform the string copy in a single pass, a typical strcpy_s function implementation retrieves a character from the source string and copies it to the destination array until it has copied the entire string or the destination array is full. If it can’t copy the entire string and s1max is positive, the strcpy_s function sets the first byte of the destination array to the null character, creating an empty string. Runtime Constraints Runtime constraints are violations of a function’s runtime requirements that the function will detect and diagnose by a call to a handler. If this handler returns, the functions will return a failure indicator to the caller. The bounds-checking interfaces enforce runtime constraints by invoking a runtime-constraint handler, which may simply return. Alternatively, the runtime-constraint handler might print a message to stderr and/or abort the program. You can control which handler function is called via the set_constraint_handler_s function, and can make the handler simply return as follows: int main(void) { constraint_handler_t oconstraint = set_constraint_handler_s(ignore_handler_s); get_y_or_n(); }

If the handler returns, the function that identified the runtime-­ constraint violation and invoked the handler indicates a failure to its caller by using its return value. The bounds-checking interface functions typically check the conditions immediately upon entry, or as they perform their tasks and gather sufficient information to determine whether a runtime constraint has been violated. The runtime constraints of the bounds-checking interfaces are conditions that would otherwise be undefined behavior for C Standard Library functions. Implementations have a default constraint handler that they invoke if no calls to the set_constraint_handler_s function have been made. The default handler’s behavior may cause the program to exit or abort, but implementations are encouraged to provide reasonable behavior by default. This allows, for example, compilers customarily used to implement safetycritical systems to not abort by default. You must check the return value of calls to functions that can return and not simply assume their results are valid. Implementation-defined behavior can be eliminated by invoking the set_constraint_handler_s function before invoking any bounds-checking interfaces or using any mechanism that invokes a runtime-­constraint handler. Annex K provides the abort_handler_s and ignore_handler_s functions, which represent two common strategies for handling errors. The C implementation’s default handler need not be either of these handlers.

Characters and Strings   143

POSIX POSIX also defines several string-handling functions, such as strdup and strndup (IEEE Std 1003.1:2018), that provide another set of stringrelated APIs for POSIX-compliant platforms. The C Standards committee published these interfaces as Technical Report 24731-2 (ISO/IEC TR 24731-2:2010), although these interfaces have not yet been included in the C Standard. These replacement functions use dynamically allocated memory to ensure that buffer overflows don’t occur, and they implement a callee allocates, caller frees model. Each function ensures that enough memory is available (except when a call to malloc fails). The strdup function, for example, returns a pointer to a new string that contains a duplicate of the argument. The returned pointer should be passed to the C Standard free function to reclaim the storage when it’s no longer needed. Listing 7-8 contains a code snippet that uses the strdup function to make a copy of the string returned by the getenv function. const char *temp = getenv("TMP"); if (temp != NULL) { char *tmpvar = strdup(temp); if (tmpvar != NULL) { printf("TMP = %s.\n", tmpvar); free(tmpvar); } } Listing 7-8: Copying a string by using the strdup function

The C Standard Library getenv function searches an environment list, provided by the host environment, for a string that matches the string referenced by a specified name ("TMP" in this example). Strings in this environment list are referred to as environment variables and provide an additional mechanism for communicating strings to a process. These strings don’t have a well-defined encoding but typically match the system encoding used for command line arguments, stdin, and stdout. The returned string (the value of the variable) may be overwritten by a subsequent call to the getenv function, so it’s a good idea to retrieve any environmental variable you need before creating any threads to eliminate the possibility of a race condition. If later use is anticipated, you should copy the string so the copy can be safely referenced as needed, as illustrated by the idiomatic example shown in Listing 7-8. The strndup function is equivalent to strdup, except that strndup copies at most n plus 1 bytes into the newly allocated memory (while strdup copies the entire string) and ensures that the newly created string is always properly terminated. These POSIX functions can help prevent buffer overflows by automatically allocating storage for the resulting strings, but this requires introducing additional calls to free when this storage is no longer needed. This

144   Chapter 7

means matching a call to free to each call to strdup or strndup, for example, which can be confusing to programmers who are more familiar with the behavior of the string functions defined by .

Microsoft Microsoft implements most of the C Standard Library functions, as well as parts of the POSIX standard. However, sometimes the Microsoft implementation of these APIs differs from the requirements of a given standard or has a function name that conflicts with an identifier reservation in another standard. In these circumstances, Microsoft will often prefix the function name with an underscore. For instance, the POSIX function strdup isn’t available on Windows, but the function _strdup is available and behaves the same way. The Visual C++ library includes the prototype implementation of the bounds-checking interfaces. Unfortunately, Visual C++ does not conform to C11 or TR 24731-1, because Microsoft chose not to update its implementation based on changes to the APIs that occurred during the standardization process. For example, Visual C++ doesn’t provide the set_constraint_handler_s function but instead retains an older function with similar behavior but an incompatible signature: _invalid_parameter_handler _set_invalid_parameter_handler(_invalid_parameter_handler)

Microsoft also doesn’t define the abort_handler_s and ignore_handler_s functions, the memset_s function (which was not defined by TR 24731-1), or the RSIZE_MAX macro. Visual C++ also doesn’t treat overlapping source and destination sequences as runtime-constraint violations and instead has undefined behavior in such cases. My NCC Group whitepaper “BoundsChecking Interfaces: Field Experience and Future Directions” provides additional information on all aspects of the bounds-checked interfaces, including Microsoft’s implementation (Seacord 2019).

Summary In this chapter, you learned about character encodings, such as ASCII and Unicode. You also learned about the various data types used to represent characters in C language programs, such as char, int, wchar_t, and so forth. We then covered character conversion libraries, including C Standard Library functions, libiconv, and Windows APIs. In addition to characters, you also learned about strings and the legacy functions and bounds-checked interfaces defined in the C Standard Library for handling strings, as well as some POSIX- and Microsoft-specific functions. In the next chapter, you’ll learn about input/output.

Characters and Strings   145

8 I N P U T/O U T P U T

This chapter will teach you how to perform input/output (I/O) operations to read data from, or write data to, terminals and filesystems. I/O involves all the ways information enters or exits a program, without which your programs would be useless. We’ll cover techniques that use C Standard streams and POSIX file descriptors. We’ll start by discussing C Standard text and binary streams. We’ll then cover different ways of opening and closing files using C Standard Library and POSIX functions. Next we’ll discuss reading and writing characters and lines, reading and writing formatted text, and reading to and from binary streams. We’ll also cover stream buffering, stream orientation, and file positioning. Many other devices and I/O interfaces (such as ioctl) are available but are not discussed in this chapter.

Standard I/O Streams The C Standard specifies the use of streams to communicate with terminals and files stored on supported, structured storage devices. A stream is a uniform abstraction for communicating with files and devices that consume or produce sequential data such as sockets, keyboards, USB ports, and printers. C uses the opaque FILE data type to represent streams. A FILE object holds the internal state information for the connection to the associated file, including the file position indicator, buffering information, an error indicator, and an end-of-file indicator. You should never allocate a FILE object yourself. C Standard Library functions operate on objects of type FILE * (that is, a pointer to the FILE type). As a result, streams are frequently referred to as file pointers. The C Standard provides an extensive API, found in , for operating on streams; we’ll explore this API later in this chapter. However, because these I/O functions need to work with a wide variety of devices and filesystems across many platforms, they’re highly abstracted, which makes them unsuitable for anything beyond the simplest applications. For example, the C Standard has no concept of directories, because it must be able to work with nonhierarchical filesystems. The C Standard makes few references to filesystem-specific details, like file permissions or locking. However, function specifications frequently state that certain behaviors happen “to the extent that the underlying system supports it,” meaning that they will occur only if they’re supported by your implementation. As a result, you’ll generally need to use the less portable APIs provided by POSIX, Windows, and other platforms to perform I/O in real-world applications. Frequently, applications will define their own APIs that, in turn, rely on platform-specific APIs to provide safe and secure I/O operations across platforms.

Stream Buffering Buffering is the process of temporarily storing data in main memory that’s passing between a process and a device or file. Buffering improves the throughput of I/O operations, which often have high latencies. Similarly, when a program requests to write to block-oriented devices like disks, the driver can cache the data in memory until it has accumulated enough data for one or more device blocks, at which point it writes the data all at once to the disk, improving throughput. This strategy is called flushing the output buffer. Like device drivers, streams often maintain their own I/O buffers. Typically, a stream uses one input buffer for each file that the program wants to read from, and one output buffer for each it wants to write to. A stream can be in one of three states: •

148   Chapter 8

Unbuffered   Characters are intended to appear from the source or at the destination as soon as possible. Streams used for error reporting or logging might be unbuffered.





Fully buffered  Characters are intended to be transmitted to or from the host environment as a block when a buffer is filled. Streams used for file I/O are normally fully buffered to optimize throughput. Line buffered  Characters are intended to be transmitted to or from the host environment as a block when a newline character is encountered. Streams connected to interactive devices such as terminals are line-buffered when you open them.

In the next section, we’ll introduce predefined streams and describe how they’re buffered.

Predefined Streams Your program has three predefined text streams open and available for use on startup. These predefined streams are declared in : extern FILE * stdin; // standard input stream extern FILE * stdout; // standard output stream extern FILE * stderr; // standard error stream

The standard output stream (stdout) is the conventional output destination from the program. This stream is usually associated with the terminal that initiated the program but can be redirected to output to a file or other stream, as follows: $ echo fred fred $ echo fred > tempfile $ cat tempfile fred

Here, the output from the echo command is redirected to tempfile. The standard input stream (stdin) is the conventional input source for the program. By default, stdin is associated with the keyboard but may be redirected to input from a file, for example, with the following commands: $ echo "one two three four five six seven" > fred $ wc < fred 1 7 34

The contents of the file fred are redirected to the stdin of the wc command, which outputs the newline (1), word (7), and byte counts (34) from fred. The standard error stream (stderr) is for writing diagnostic output. As initially opened, stderr isn’t fully buffered; stdin and stdout are fully buffered if and only if the stream doesn’t refer to an interactive device. The stderr stream isn’t fully buffered, so that error messages can be viewed as soon as possible. Figure 8-1 shows the predefined streams stdin, stdout, and stderr attached to the keyboard and display of the user’s terminal. Input/Output   149

stdout

Process

stderr stdin

Figure 8-1: Standard streams attached to I/O communication channels

The output stream of one program can be redirected to be another application’s input stream by using POSIX pipes. In many operating systems, you can chain applications together by separating commands with the vertical bar character (|): $ echo "Hello Robert" | sed "s/Hello/Hi/" | sed "s/Robert/robot/" Hi robot

Stream Orientation Each stream has an orientation that indicates whether the stream contains narrow or wide characters. After a stream is associated with an external file, but before any operations are performed on it, the stream doesn’t have an orientation. Once a wide character I/O function has been applied to a stream without orientation, the stream becomes a wide-oriented stream. Similarly, once a byte I/O function has been applied to a stream without orientation, the stream becomes a byte-oriented stream. Multibyte character sequences or narrow characters that can be represented as an object of type char (that are required by the C Standard to be 1 byte) can be written to a byte-oriented stream. You can reset the orientation of a stream by using the fwide function or by closing and then reopening the file. Applying a byte I/O function to a wide-oriented stream, or a wide-character I/O function to a byte-oriented stream, results in undefined behavior. You should never mix narrow character data, wide character data, and binary data in the same file. All three predefined streams (stderr, stdin, and stdout) are unoriented at program startup.

Text and Binary Streams The C Standard supports both text streams and binary streams. A text stream is an ordered sequence of characters composed into lines, each of which consists of zero or more characters plus a terminating newline character sequence. You can denote a single line break on a Unix-like system by using a line feed (\n). Most Microsoft Windows programs use a carriage return (\r) followed by a line feed (\n). The different newline conventions cause text files that have been transferred between systems with different conventions to display incorrectly. Text in files created with programs that are common on Unix-like 150   Chapter 8

operating systems appear as a single, long line on older Microsoft Windows programs that don’t display a single line feed or a single carriage return as a line break. A binary stream is an ordered sequence of arbitrary binary data. Data read in from a binary stream will be the same as data written out earlier to that same stream, under the same implementation. These streams may, however, have an implementation-defined number of null bytes appended to the end of the stream. Binary streams are always more capable and more predictable than a text stream. However, the easiest way to read or write an ordinary text file that can work with other text-oriented programs is through a text stream.

Opening and Creating Files When you open or create a file, it’s associated with a stream. The following functions open or create a file.

The fopen Function The fopen function opens the file whose name is given as a string and pointed to by filename, and then associates a stream with it. If the file doesn’t already exist, fopen will create it: FILE *fopen( const char * restrict filename, const char * restrict mode );

The mode argument points to one of the strings shown in Table 8-1 to determine how to open the file. Table 8-1: Valid File Mode Strings

Mode string

Description

r

Open existing text file for reading

w

Truncate to zero length or create text file for writing

a

Append, open, or create text file for writing at end-of-file

rb

Open existing binary file for reading

wb

Truncate file to zero length or create binary file for writing

ab

Append, open, or create binary file for writing at end-of-file

r+

Open existing text file for reading and writing

w+

Truncate to zero length or create text file for reading and writing

a+

Append; open or create text file for update, writing at current end-of-file

r+b or rb+

Open existing binary file for update (reading and writing)

w+b or wb+

Truncate to zero length or create binary file for reading and writing

a+b or ab+

Append; open or create binary file for update, writing at current end-of-file Input/Output   151

Opening a file with read mode (by passing r as the first character in the mode argument) fails if the file doesn’t exist or cannot be read. Opening a file with append mode (by passing a as the first character in the mode argument) causes all subsequent writes to the file to occur at the current end of the file. In some implementations, opening a binary file with append mode (by passing b as the second or third character in the mode argument) may initially set the file position indicator for the stream beyond the last data written, because of null character padding. You can open a file in update mode by passing + as the second or third character in the mode argument, allowing both read and write operations to be performed on the associated stream. Opening (or creating) a text file with update mode may instead open (or create) a binary stream in some implementations. The C11 Standard added the exclusive mode for reading and writing binary and text files, as shown in Table 8-2. Table 8-2: Valid File Mode Strings Added by C11

Mode string

Description

wx

Create exclusive text file for writing

wbx

Create exclusive binary file for writing

w+x

Create exclusive text file for reading and writing

w+bx or wb+x

Create exclusive binary file for reading and writing

Opening a file with exclusive mode (by passing x as the last character in the mode argument) fails if the file already exists or cannot be created. Otherwise, the file is created with exclusive (also known as nonshared) access to the extent that the underlying system supports exclusive access. On a final note, make sure that you never copy a FILE object. The following program, for example, can fail because a by-value copy of stdout is being used in the call to fputs: #include #include int main(void) { FILE my_stdout = *stdout; if (fputs("Hello, World!\n", &my_stdout) == EOF) { return EXIT_FAILURE; } return EXIT_SUCCESS; }

This program has undefined behavior and it typically results in a crash when executed.

152   Chapter 8

The POSIX open Function On POSIX systems, the open function (IEEE Std 1003.1:2018) establishes the connection between a file identified by path and a value called a file descriptor: int open(const char *path, int oflag, ...);

The file descriptor is a non-negative integer that refers to the structure representing the file (called the open file description). The file descriptor returned by the open function is the smallest integer not yet returned by a prior call to open or passed to close and is unique to the calling process. The file descriptor is used by other I/O functions to refer to that file. The open function sets the file offset used to mark the current position within the file to the beginning of the file. The value of the oflag parameter sets the open file description’s file access modes, which specify whether the file is being opened for reading, writing, or both. Values for oflag are constructed by a bitwise-inclusive OR of a file access mode and any combination of access flags. Applications must specify exactly one of the following file access modes in the value of oflag: O_EXEC  Open for execute only (nondirectory files) O_RDONLY  Open for reading only O_RDWR  Open for reading and writing O_SEARCH  Open directory for search only O_WRONLY  Open for writing only

The value of the oflag parameter also sets the file status flags, which control the behavior of the open function and affect how file operations are performed. These flags include the following: O_APPEND  Sets the file offset to the end of the file prior to each write O_TRUNC  Truncates the length to 0 O_CREAT  Creates a file O_EXCL  Causes the open to fail if O_CREAT is also set and the file exists

The open function takes a variable number of arguments. The value of the argument following the oflag argument specifies the file-mode bits (the file permissions when you create a new file) and is of type mode_t. Listing 8-1 shows an example of using the open function to open a file for writing by the owner. #include #include #include #include //---snip--int fd; u mode_t mode = S_IRUSR | S_IWUSR | S_IRGRP | S_IROTH; const char *pathname = "/tmp/file"; Input/Output   153

//---snip--if ((fd = open(pathname, O_WRONLY | O_CREAT | O_TRUNC, mode)v) == -1) { fprintf(stderr, "Can’t open %s.\n", pathname); exit(1); } //---snip--Listing 8-1: Opening a file as write-only by the owner

We create a mode flag u that is a bitwise-inclusive OR of the following mode bits for access permission: S_IRUSR  Read permission bit for the owner of the file S_IWUSR  Write permission bit for the owner of the file S_IRGRP  Read permission bit for the group owner of the file S_IROTH  Read permission bit for other users

The call to open v takes multiple arguments, including the pathname of the file, the oflag, and the mode. The file access mode is O_WRONLY, which means the file is opened for writing only. The O_CREAT file status flag informs open to create the file; the O_TRUNC file status flag informs open that if the file exists and is successfully opened, it should discard the previous contents of the file but retain the identifier. If the file was successfully opened, the open function returns a nonnegative integer representing the file descriptor. Otherwise, open returns −1 and sets errno to indicate the error. Listing 8-1 checks for a value of −1, writes a diagnostic message to the predefined stderr stream if an error occurred, and then exits. In addition to open, POSIX has other useful functions for working with file descriptors, such as the fileno function to get the file descriptor associated with an existing file pointer, and the fdopen function to create a new stream file pointer from an existing file descriptor. POSIX APIs available through the file descriptor allow access to features of POSIX filesystems that aren’t normally exposed through the file pointer interfaces such as directories, file permissions, and symbolic and hard links.

Closing Files Opening a file allocates resources. If you continually open files without closing them, you’ll eventually run out of file descriptors or handles available for your process, and attempting to open more files will fail. Consequently, it’s important to close files after you’ve finished using them.

The fclose Function The C Standard Library fclose function closes the file: int fclose(FILE *stream);

154   Chapter 8

Any unwritten buffered data for the stream is delivered to the host environment to be written to the file. Any unread buffered data is discarded. It’s possible for the fclose function to fail. When fclose writes the remaining buffered output, for example, it might return an error because the disk is full. Even if you know the buffer is empty, errors can still occur when closing a file if you’re using the Network File System (NFS) protocol. Despite the possibility of failure, recovery is often impossible, so programmers commonly ignore errors returned by fclose. When closing the file fails, a common practice is to abort the process or to truncate the file so its contents are meaningful when next read. The value of a pointer to a FILE object is indeterminate after the associated file is closed. Whether a file of zero length (in which an output stream hasn’t written any data) actually exists is implementation defined. You can reopen a closed file in the same program or another one, and its contents can be reclaimed or modified. If the main function returns, or if the exit function is called, all open files close (and all output streams are flushed) before program termination. Other paths to program termination, such as calling the abort function, may not close all files properly, which means that buffered data not yet written to a disk might be lost.

The POSIX close Function On POSIX systems, you can use the close function to deallocate the file descriptor specified by fildes: int close(int fildes);

If an I/O error occurred while reading from or writing to the filesystem during close, it may return −1 with errno set to EIO; if this error is returned, the state of fildes is unspecified, meaning you can no longer read or write data to the descriptor or attempt to close it again. Once a file is closed, the file descriptor no longer exists, because the integer corresponding to it no longer refers to a file. Files are also closed when the process owning that stream terminates. An application that uses fopen to open a file must use fclose to close it; an application that uses open to open a file must use close to close it (unless it passed the descriptor to fdopen, in which case it must close by calling fclose).

Reading and Writing Characters and Lines The C Standard defines functions for reading and writing specific characters or lines. Most byte stream functions have counterparts that take a wide character (wchar_t) or wide character string instead of a narrow character (char)

Input/Output   155

or string, respectively (see Table 8-3). Byte stream functions are declared in the header file , and the wide stream functions in . The wide-character functions operate on the same streams (such as stdout). Table 8-3: Narrow and Wide String I/O Functions char

wchar_t

Description

fgetc

fgetwc

Reads a character from a stream

getc

getwc

Reads a character from a stream (often a macro)

getchar

getwchar

Reads a character from stdin

fgets

fgetws

Reads a line from a stream

fputc

fputwc

Writes a character to a stream

putc

putwc

Writes a character to a stream (often a macro)

fputs

fputws

Writes a string to a stream

putchar

putwchar

Writes a character to stdout

puts

N/A

Writes a string to stdout

ungetc

ungetwc

Returns a character to a stream

scanf

wscanf

Reads formatted character input from stdin

fscanf

fwscanf

Reads formatted character input from a stream

sscanf

swscanf

Reads formatted character input from a buffer

printf

wprintf

Prints formatted character output to stdout

fprintf

fwprintf

Prints formatted character output to a stream

sprintf

swprintf

Prints formatted character output to a buffer

snprintf

N/A

The same as sprintf with truncation. The swprintf function also takes a length argument, but behaves differently from snprintf in the way it interprets it.

In this chapter, we’ll discuss the byte stream functions only. You may want to avoid wide-character function variants altogether and work exclusively with UTF-8 character encodings, if possible, as these functions are less prone to programmer error and security vulnerabilities. The fputc function converts the character c to the type unsigned char and writes it to stream: int fputc(int c, FILE *stream);

It returns EOF if a write error occurs; otherwise, it returns the character it has written. The putc function is just like fputc, except that most implementations implement it as a macro: int putc(int c, FILE *stream);

If putc is implemented as a macro, it may evaluate its arguments more than once, so the arguments should never be expressions with side effects. 156   Chapter 8

Using fputc is generally safer. See CERT C rule FIO41-C (Do not call getc(), putc(), getwc(), or putwc() with a stream argument that has side effects) for more information. The putchar function is equivalent to the putc function, except that it uses stdout as the value of the stream argument. The fputs function writes the string s to the stream stream: int fputs(const char * restrict s, FILE * restrict stream);

This function doesn’t write the null character from the string s—nor does it write a newline character—but outputs only the characters in the string. If a write error occurs, fputs returns EOF. Otherwise, it returns a nonnegative value. For example, the following statements output the text I am Groot, followed by a newline: fputs("I ", stdout); fputs("am ", stdout); fputs("Groot\n", stdout);

The puts function writes the string s to the stream stdout followed by a newline: int puts(const char *s);

The puts function is the most convenient function for printing simple messages because it takes only a single argument. Here’s an example: puts("This is a message.");

The fgetc function reads the next character as an unsigned char from a stream and returns its value, converted to an int: int fgetc(FILE *stream);

If an end-of-file condition or read error occurs, the function returns EOF. You may recall that the gets function reads characters from stdin and writes them into a character array until a newline or EOF is reached. The gets function is inherently insecure. It was deprecated in C99 and removed from C11 and should never be used. If you need to read a string from stdin, consider using the fgets function instead. The fgets function reads at most one less than the number of characters specified (leaving room for a null character) from a stream into a character array.

Stream Flushing As described earlier in this chapter, streams can be fully or partially buffered, meaning that data you thought you wrote may not yet be delivered to the host environment. In particular, this can be a problem when the Input/Output   157

program terminates abruptly. The fflush function delivers any unwritten data for a specified stream to the host environment to be written to the file: int fflush(FILE *stream);

The behavior is undefined if the last operation on the stream was input. If the stream is a null pointer, the fflush function performs this flushing action on all streams. You should make sure that your file pointer isn’t null before calling fflush if this isn’t your intent. The fflush function sets the error indicator for the stream and returns EOF if a write error occurs; otherwise, it returns zero.

Setting the Position in a File Random-access files (which include a disk file, for example, but not a terminal) maintain a file position indicator associated with the stream. The file position indicator describes where in the file the stream is currently reading or writing. When you open a file, the indicator is positioned at the file’s start. You can position the indicator whenever you wish to read or write any portion of the file. The ftell function obtains the current value of the file position indicator, while the fseek function sets the file position indicator. These functions use the long int type to represent offsets (positions) in a file and are therefore limited to offsets that can be represented as a long int. Listing 8-2 demonstrates the use of the ftell and fseek functions. #include #include int main(void) { FILE *fp = fopen("fred.txt", "r"); if (fp == NULL) { fputs("Cannot open fred.txt file\n", stderr); return EXIT_FAILURE; } if (fseek(fp, 0, SEEK_END) != 0) { fputs("Seek to end of file failed\n", stderr); return EXIT_FAILURE; } long int fpi = ftell(fp); if (fpi == -1L) { perror("Tell"); return EXIT_FAILURE; } printf("file position = %ld\n", fpi); if (fclose(fp) == EOF) { fputs("Failed to close file\n", stderr); return EXIT_FAILURE; }

158   Chapter 8

return EXIT_SUCCESS; } Listing 8-2: Using the ftell and fseek functions

This program opens a file called fred.txt and calls fseek to set the file position indicator to the end of the file (indicated by SEEK_END). The ftell function returns the current value of the file position indicator for the stream as a long int. The program prints out this value and exits. Finally, we close the file referenced by the fp file pointer. To ensure your code is robust, make sure you check for errors. File I/O, in particular, can fail for any number of reasons. The fopen function returns a null pointer when it fails. The fseek function returns nonzero only for a request that cannot be satisfied. On failure, the ftell function returns −1L and stores an implementation-defined value in errno. If the return value from ftell is equal to −1L, we use the perror function to print the string we provided “Tell” followed by a colon (:), a space, an appropriate error message corresponding to the value stored in errno, and finally, a newline character. The fclose function returns EOF if any errors were detected. One of the unfortunate aspects of the C Standard Library demonstrated by this short program is that each function tends to report errors in its own unique way, so you normally need to refer to your documentation to see how to test for errors. The newer fgetpos and fsetpos functions use the fpos_t type to represent offsets. This type can represent arbitrarily large offsets, meaning you can use fgetpos and fsetpos with arbitrarily large files. A wide-oriented stream has an associated mbstate_t object that stores the stream’s current parse state. A successful call to fgetpos stores this multibyte state information as part of the value of the fpos_t object. A later successful call to fsetpos using the same stored fpos_t value restores the parse state as well as the position within the controlled stream. It’s not possible to convert an fpos_t object to an integer byte or character offset within the stream except indirectly by calling fsetpos followed by ftell. The short program shown in Listing 8-3 demonstrates the use of the fgetpos and fsetpos functions. #include #include int main(void) { FILE *fp = fopen("fred.txt", "w+"); if (fp == NULL) { fputs("Cannot open fred.txt file\n", stderr); return EXIT_FAILURE; } fpos_t pos; if (fgetpos(fp, &pos) != 0) { perror("get position"); return EXIT_FAILURE; } if (fputs("abcdefghijklmnopqrstuvwxyz", fp) == EOF) { fputs("Cannot write to fred.txt file\n", stderr); } Input/Output   159

if (fsetpos(fp, &pos) != 0) { perror("set position"); return EXIT_FAILURE; } long int fpi = ftell(fp); if (fpi == -1L) { perror("seek"); return EXIT_FAILURE; } printf("file position = %ld\n", fpi); if (fputs("0123456789", fp) == EOF) { fputs("Cannot write to fred.txt file\n", stderr); } if (fclose(fp) == EOF) { fputs("Failed to close file\n", stderr); return EXIT_FAILURE; } return EXIT_SUCCESS; } Listing 8-3: Using the fgetpos and fsetpos functions

This program opens the fred.txt file for writing and then calls fgetpos to get the current file position within the file, which is stored in pos. We then write some text to the file before calling fsetpos to restore the file position indicator to the position stored in pos. At this point, we can use the ftell function to retrieve and print the file position, which should be 0. After running this program, fred.txt contains the following text: 0123456789klmnopqrstuvwxyz

You cannot write to a stream and then read from it again without an intervening call to the fflush function to write any unwritten data or to a file positioning function (fseek, fsetpos, or rewind). You also cannot read from a stream and then write to it without an intervening call to a file positioning function. The rewind function sets the file position indicator to the beginning of the file: void rewind(FILE *stream);

The rewind function is equivalent to the invoking fseek followed by clearerr to clear the error indicator for the stream: fseek(stream, 0L, SEEK_SET); clearerr(stream);

You shouldn’t attempt to use file positions in files opened in append mode, because many systems don’t modify the current file position indicator for append or will forcefully reset to the end of the file when writing. If using the APIs that use file positions, the file position indicator is 160   Chapter 8

maintained by subsequent reads, writes, and positioning requests. Both POSIX and Windows have APIs that never use the file position indicator; for those, you always specify the offset into the file at which to perform I/O.

Removing and Renaming Files The C Standard Library provides a remove function to delete a file, and a rename function to move or rename it: int remove(const char *filename); int rename(const char *old, const char *new);

In POSIX, the file deletion function is named unlink: int unlink(const char *path);

The unlink function has better-defined semantics because it’s specific to POSIX filesystems. POSIX also uses rename for renaming. In POSIX and Windows, we can have any number of links to a file, including hard links and open file descriptors. The unlink function always removes the directory entry for the file, but it deletes the actual file only when there are no more links to it. Even at this point, the actual contents of the file may remain in permanent storage. The remove function behaves the same as the unlink function on POSIX systems but may have different behavior on other operating systems.

Using Temporary Files We frequently use temporary files as an interprocess communication mechanism or for temporarily storing information out to disk to free up RAM. For example, one process might write to a temporary file that another process reads from. These files are normally created in a temporary directory by using functions such as the C Standard Library’s tmpfile and tmpnam or POSIX’s mkstemp. Temporary directories can be either global or user specific. In Unix and Linux, the TMPDIR environment variable is used to specify the location of the global temporary directories, which are typically /tmp and /var/tmp. Linux usually has user-specific temporary directories defined by the $XDG_RUNTIME_DIR environment variable, which is typically set to /run/user/$uid. In Windows, you can find user-specific temporary directories in the AppData section of the User Profile, typically C:\Users\User Name\AppData\Local\Temp (%USERPROFILE%\AppData\Local\Temp). On Windows, the global temporary directory is specified by either the TMP, TEMP, or USERPROFILE environment variable. The C:\Windows\Temp directory is a system folder used by Windows to store temporary files. For security reasons, it’s best for each user to have their own temporary directory, because the use of global temporary directories frequently results Input/Output   161

in security vulnerabilities. The most secure function for creating temporary files is the POSIX mkstemp function. However, because accessing files in shared directories may be difficult or impossible to implement securely, we recommended that you not use any of the available functions, and instead perform the interprocess communication by using sockets, shared memory, or other mechanisms designed for this purpose.

Reading Formatted Text Streams In this section, we’ll demonstrate the use of the fscanf function to read formatted input. The fscanf function is the corresponding input version of the fprintf function that we introduced all the way back in Chapter 1 and has the following signature: int fscanf(FILE * restrict stream, const char * restrict format, ...);

The fscanf function reads input from the stream pointed to by stream, under control of the format string that tells the function how many arguments to expect, their type, and how to convert them for assignment. Subsequent arguments are pointers to the objects receiving the converted input. If you provide fewer arguments than conversion specifiers, the behavior is undefined. If you provide more arguments than conversion specifiers, the excess arguments are evaluated but otherwise ignored. The fscanf function has a great deal of functionality that we’ll only touch upon here. For more information, refer to the C Standard. To demonstrate the use of fscanf, as well as some other I/O functions, we’ll implement a program that reads in the signals.txt file shown in List­ ing 8-4 and print each line. Each line of this file contains the following: • • •

A signal number (a small, positive integer value) The signal ID (a small string of up to six alphanumeric characters) A short string with a description of the signal

Fields are whitespace delimited except for the description, which can contain whitespace and is delimited by a newline: 1 2 3 4 5 6 7 8

HUP Hangup INT Interrupt QUIT Quit ILL Illegal instruction TRAP Trace trap ABRT Abort EMT EMT trap FPE Floating-point exception

Listing 8-4: The signals.txt file

162   Chapter 8

Listing 8-5 shows the signals program, which reads this file and prints out each line. #include #include #include int main(void) { int status = EXIT_SUCCESS; FILE *in; struct sigrecord { int signum; char signame[10]; char sigdesc[100]; u } sigrec; if ((in = fopen("signals.txt", "r")) == NULL) { fputs("Cannot open signals.txt file\n", stderr); return EXIT_FAILURE; } do { v int n = fscanf(in, "%d%9s%*[ \t]%99[^\n]", &sigrec.signum, sigrec.signame, sigrec.sigdesc ); if (n == 3) { printf( "Signal\n number = %d\n name = %s\n description = %s\n\n", sigrec.signum, sigrec.signame, sigrec.sigdesc ); } else if (n != EOF) { fputs("Failed to match signum, signame or sigdesc\n", stderr); status = EXIT_FAILURE; break; } else { break; } w } while (1); x if (fclose(in) == EOF) { fputs("Failed to close file\n", stderr); status = EXIT_FAILURE; } return status; } Listing 8-5: The signals program

We define several variables in the main function, including the ­sigrec structure u, which we’ll use to store the signal information found on each line of the file. The sigrec structure contains three members: a signum member of Input/Output   163

type int that will hold the signal number; a signame member that’s an array of char and will hold the signal ID; and the sigdesc member, an array of char that will hold the description of the signal. Both arrays have fixed sizes that we determined were adequately sized for the strings being read from the file. If the strings read from the file are too long to fit in these arrays, the program will treat it as an error. The call to fscanf v reads each line of input from the file. It appears inside of an infinite do...while (1) loop w that we must break out for the program to terminate. We assign the return value from the fscanf function to a local variable n. The fscanf function returns EOF if an input failure occurs before the first conversion has completed. Otherwise, the function returns the number of input items assigned, which can be fewer than provided for, or even zero, in the event of an early matching failure. The call to fscanf assigns three input items, so we only print the signal description when n is equal to 3. Otherwise, if n does not equal EOF, a matching failure occurred, so we output an appropriate diagnostic message to stderr, set status to EXIT_FAILURE, and break out of the loop. The final possibility is that fscanf returns EOF indicating that the end of the file was reached, in which case we simply break out of the loop without altering status. The fscanf function uses a format string to inform it how to assign the input text to each argument. In this case, the "%d%9s%*[ \t]%99[^\n]" format string contains four conversion specifiers, which specify how the input read from the stream is converted into values stored in the objects referenced by the format string’s arguments. We introduce each conversion specification with the character %. After the %, the following may appear, in sequence: • • • •

An optional character * that discards the input without assigning it to an argument An optional integer greater than zero that specifies the maximum field width (in characters) An optional length modifier that specifies the size of the object A conversion specifier character that specifies the type of conversion to be applied

The first conversion specifier in the format string is %d. This conversion specifier matches the first optionally signed decimal integer—which should correspond to the signal number in the file—and stores the value in the third argument referenced by sigrec.signum. Without an optional length modifier, the length of the input depends on the conversion specifier’s default type. For the d conversion specifier, the argument must point to a signed int. The second conversion specifier in this format string is %9s, which matches the next sequence of nonwhitespace characters from the input stream—corresponding to the signal name—and stores these characters as a string in the fourth argument referenced by sigrec.signame. The length modifier prevents more than nine characters from being input then writes a null character in sigrec.signame after the characters that have been matched. A conversion specifier of %10s in this example would allow a buffer 164   Chapter 8

overflow to occur. A conversion specifier of %9s can still fail to read the entire string, resulting in a matching error. When reading data into a fixedsize buffer as we are doing, it is a good idea to test inputs that exactly match or slightly exceed the fixed buffer length to ensure buffer overflow does not occur and that the string is properly null-terminated. We’re going to skip the third conversion specifier for a moment and talk about the fourth one: %99[^\n]. This fancy conversion specifier will match the signal description field in the file. The brackets ([]) contain a scanset, which is similar to a regular expression. This scanset uses the circumflex (^) to exclude \n characters. Put together, %99[^\n] reads all the characters until it reaches a \n (or EOF), and stores them in the fifth argument referenced by sigrec.sigdesc. C programmers commonly use this syntax to read an entire line. This conversion specifier also includes a maximum string length of 99 characters to avoid buffer overflows. Finally, we’ll circle back to the third conversion specifier: %*[ \t]. As we have just seen, the fourth conversion specifier reads all the characters, starting from the end of the signal ID. Unfortunately, this includes any whitespace characters between the signal ID and the start of the description. The purpose of the %*[ \t] conversion specifier is to consume any space or horizontal tab characters between these two fields and suppress them by using the assignment-suppressing specifier *. It’s also possible to include other whitespace characters in the scanset for this conversion specifier. Finally, we close the file by calling the fclose function x.

Reading to and Writing from Binary Streams The C Standard Library fread and fwrite functions operate on a binary stream. The fwrite function has the following signature: size_t fwrite(const void * restrict ptr, size_t size, size_t nmemb, FILE * restrict stream);

This function writes up to nmemb elements of size bytes from the array pointed to by ptr to stream. It does so by converting each object to an array of unsigned char (every object can be converted to an array of this type), and then calling the fputc function to write the value of each character in the array in order. The file position indicator for the stream is advanced by the number of characters successfully written. Listing 8-6 demonstrates the use of the fwrite function to write signal records to the signals.txt file. #include #include #include typedef struct sigrecord { int signum; char signame[10]; Input/Output   165

char sigdesc[100]; } sigrecord; int main(void) { int status = EXIT_SUCCESS; FILE *fp; sigrecord sigrec; u if ((fp = fopen("signals.txt", "wb")) == NULL) { fputs("Cannot open signals.txt file\n", stderr); return EXIT_FAILURE; } v sigrecord sigrec30 = { 30, "USR1", "user-defined signal 1" }; sigrecord sigrec31 = { .signum = 31, .signame = "USR2", .sigdesc = "user-defined signal 2" }; size_t size = sizeof(sigrecord); w if (fwrite(&sigrec30, size, 1, fp) != 1) { fputs("Cannot write sigrec30 to signals.txt file\n", stderr); status = EXIT_FAILURE; goto close_files; } if (fwrite(&sigrec31, size, 1, fp) != 1) { fputs("Cannot write sigrec31 to signals.txt file\n", stderr); status = EXIT_FAILURE; } close_files: if (fclose(fp) == EOF) { fputs("Failed to close file\n", stderr); status = EXIT_FAILURE; } return status; } Listing 8-6: Writing to a binary file using direct I/O

We open the signals.txt file in wb mode u to create a binary file for writing. We declare two sigrecord structures v and initialize them with the signal values we want to write to the file. The second structure, sigrec31, is initialized using designated initializers for comparison. Both initialization styles have the same behavior; designated initializers make the declaration less terse but clearer. The actual writing begins at w. We check the return values from each call to the fwrite function to ensure that it wrote the correct number of elements. Listing 8-7 uses the fread function to read the data we just wrote from the signals.txt file.

166   Chapter 8

#include #include #include typedef struct sigrecord { int signum; char signame[10]; char sigdesc[100]; } sigrecord; int main(void) { int status = EXIT_SUCCESS; FILE *fp; sigrecord sigrec; size_t size = sizeof(sigrecord); u if ((fp = fopen("signals.txt", "rb")) == NULL) { fputs("Cannot open signals.txt file\n", stderr); return EXIT_FAILURE; } // read the second signal v if (fseek(fp, size, SEEK_SET) != 0) { fputs("fseek in signals.txt file failed\n", stderr); status = EXIT_FAILURE; goto close_files; } w if (fread(&sigrec, size, 1, fp) != 1) { fputs("Cannot read from signals.txt file\n", stderr); status = EXIT_FAILURE; goto close_files; } printf( "Signal\n number = %d\n name = %s\n description = %s\n\n", sigrec.signum, sigrec.signame, sigrec.sigdesc ); close_files: fclose(fp); return status; } Listing 8-7: Reading from a binary file using direct I/O

We open the binary file by using the rb mode u for reading. Next, to make this example a bit more interesting, the program reads and prints the information for a specific signal, rather than reading the entire file. We could indicate which signal to read by using an argument to the program, but for this example, we hardcoded it as the second signal. To accomplish this, the program invokes the fseek function v to set the file position indicator for the stream referenced by fp. As you learned earlier in this chapter, Input/Output   167

the file position indicator determines the file position for the subsequent I/O operation. For a binary stream, we set the new position by adding the offset (measured in bytes) to the position specified by the final argument (the beginning of the file, as indicated by SEEK_SET). The first signal is at position 0 in the file, and each subsequent signal is at an integer multiple of the size of the structure from the beginning of the file. After the file position indicator is positioned at the start of the second signal, we call the fread function w to read the data from the binary file into the structure referenced by &sigrec. Similar to fwrite, the fread function reads up to one element, whose size is specified by size, from the stream pointed to by fp. In most cases, this object has the size and type of the corresponding call to fwrite. The file position indicator for the stream is advanced by the number of characters successfully read. We check the return value from the fread function to ensure the correct number of e ­ lements was read. Binary files can have different formats. In particular, the order of bytes within a binary representation of a number, or endianness, can vary between systems. A big-endian ordering places the most significant byte first and the least significant byte last, while a little-endian ordering does the opposite. For example, consider the unsigned hexadecimal number 0x1234, which requires at least two bytes to represent. In a big-endian ordering, these two bytes are 0x12, 0x34, while in a little-endian ordering, the bytes are arranged as 0x34, 0x12. Intel and AMD processors use the little-endian format, while the ARM and POWER series of processors can switch between either littleor big-endian format. However, big-endianness is the dominant ordering in network protocols such as IP, TCP, and UDP. Endianness is a problem when a binary file is created on one computer and is read on another computer with different endianness. Endianness independence in binary data ­formats can be achieved by always storing the data in one fixed endianness, or including a field in the binary file to indicate the endianness of the data.

Summary In this chapter, you learned about I/O as well as C Standard streams, including stream buffering, the predefined streams, stream orientation, and the difference between text and binary streams. You then learned how to create, open, and close files by using the C Standard Library and POSIX APIs. You also learned how to read and write characters and lines, read and write formatted text, and read and write from binary streams. You looked at how to flush a stream, set the position in a file, remove and rename files. Finally, you learned about temporary files and to avoid using them. In the next chapter, you’ll learn about compilation process and the preprocessor, including file inclusion, conditional inclusion, and macros.

168   Chapter 8

9 PREPROCESSOR with Aaron Ballman

The preprocessor is the part of the C compiler that runs at an early phase of compilation and transforms the source code before it’s translated, such as inserting code from one file (typically a header file) into another (typically a source file). The preprocessor also allows you to specify that an identifier should be automatically substituted by a source code segment during macro expansion. In this chapter, you’ll learn how to use the preprocessor to include files, define object- and function-like macros, and conditionally include code based on implementation-specific features.

The Compilation Process Let’s first see where the preprocessor sits in the compilation process. Conceptually, the compilation process consists of a pipeline of eight phases, as shown in Figure 9-1. We call these translation phases because each phase translates the code for processing by the next phase. Character   mapping

Line  splicing

Tokenization

Preprocessing

Character-set   mapping

  String concatenation

Translation

Linkage

Figure 9-1: Translation phases

The preprocessor runs before the source code is translated into object code by the translator. This allows the preprocessor to modify the source code written by the user before it’s operated on by the translator. Consequently, the preprocessor has a limited amount of semantic information about the program being compiled. It doesn’t know about functions, variables, or types. Instead, only basic elements, such as header names, identifiers, literals, and punctuation characters like +, -, and ! are meaningful to the preprocessor. These basic elements, called tokens, are the smallest elements of a computer program that have meaning to a compiler. The preprocessor operates on preprocessing directives that you include in the source code to program the behavior of the preprocessor. You spell preprocessing directives with a leading # token followed by a directive name, such as #include, #define, or #if. You can include whitespace between the beginning of the line and the # or between the # and the directive to indent directives. Terminate each preprocessing directive with a newline character. A preprocessing directive causes the preprocessor to take an action that might alter the resulting translation unit, meaning the code you write is often not the same code consumed by the translator. Compiler implementations usually provide a way to view the preprocessor output, called a translation unit, passed to the translator. Viewing the preprocessor output is unnecessary, but you may find it informative to see the actual code given to the translator. Table 9-1 lists flags used by common compilers to output a translation unit. Preprocessed output files are commonly given a .i file extension. Table 9-1: Outputting a Translation Unit

170   Chapter 9

Compiler

Example command line

Clang

clang other-options -E -o output_file.i source.c

GCC

gcc other-options -E -o output_file.i source.c

Visual C++

cl other-options /P /Fioutput_file.i source.c

File Inclusion A powerful feature of the preprocessor is the ability to insert the contents of one source file into the contents of another source file by using the #include preprocessing directive. The included files are called header files to distinguish them from other source files. Header files typically contain declarations for use by other programs. This is the most common way to share external declarations of functions and objects with other parts of the program. You’ve already seen many examples of including the headers for C Standard Library functions in the examples in this book. For instance, the program in Table 9-2 is separated into a header file named bar.h and a source file named foo.c. The source file foo.c does not directly contain a declaration of func, yet the function is successfully referenced by name within main. During preprocessing, the #include directive inserts the contents of bar.h into foo.c in the place of the #include directive itself. Table 9-2: Header File Inclusion

Original sources

Resulting translation unit

bar.h

int func(void);   int main(void) {  return func();  }

int func(void);

foo.c #include "bar.h"    int main(void) {  return func();  }

The preprocessor executes #include directives as it encounters them. Therefore, inclusion has transitive properties: if a source file includes a header file that itself includes another header file, the preprocessed output will contain the contents of both header files. For example, given the baz.h and bar.h header files and the foo.c source file, the output after running the preprocessor on the foo.c source code is shown in Table 9-3. Table 9-3: Transitive Header File Inclusion

Original sources

Resulting translation unit

baz.h

int other_func(void); 

int other_func(void);

bar.h #include "baz.h"    int func(void);

int func(void);    int main(void) {  return func();  }

foo.c #include "bar.h"    int main(void) {  return func();  } Preprocessor

   171

Compiling the foo.c source file causes the preprocessor to include the "bar.h" header file. The preprocessor then finds the include directive for the "baz.h" header file and includes it as well, bringing the declaration for other_func into the resulting translation unit.

Quoted and Angle Bracket Include Strings You can use either a quoted include string (for example, #include "foo.h") or an angle-bracketed include string (for example, #include ) to specify the file to include. The difference between these syntaxes is i­ mplementation defined, but typically influences the search path used to find the included files. For example, both Clang and GCC attempt to find files included with: • •

Angle brackets on the system include path, specified using the -isystem flag Quoted strings on the quoted include path, specified using the -iquote flag

Refer to your compiler’s documentation for the specific differences between these two syntaxes. Normally, headers for standard or system libraries are found on the default system include path, and your own pro­ ject headers are found on the quoted include path.

Conditional Inclusion Frequently, you’ll need to write different code to support different implementations. For example, you may want to provide alternative implementations of a function for different target architectures. One solution to this problem is to maintain two files with slight variations between them and compile the appropriate file for a particular implementation. A better solution is to either translate or refrain from translating the target-specific code based on a preprocessor definition. You can conditionally include source code by including a predicate condition using the preprocessing directives #if, #elif, and #else. A predicate condition is the controlling constant expression that’s evaluated to determine which branch of the program the preprocessor should take. They’re typically used along with the preprocessor defined operator, which determines if a given identifier is the name of a defined macro. The conditional inclusion directives are similar to the if and else statements. When the predicate condition evaluates to a nonzero preprocessor value, the #if branch is processed, and all other branches are not. When the predicate condition evaluates to zero, the next #elif branch, if any, has its predicate tested for inclusion. If none of the predicate conditions evaluate to nonzero, then the #else branch, if there is one, is processed. The #endif preprocessing directive indicates the end of the conditionally included code.

172   Chapter 9

The defined operator evaluates to 1 if the given identifier is defined as a macro, or 0 otherwise. For example, the preprocessing directives shown in Listing 9-1 conditionally determine which header file contents to include in the resulting translation unit. The preprocessed output of foo.c depends on whether _WIN32 or __ANDROID__ is a defined macro. If neither is a defined macro, the preprocessor output will be empty. /* foo.c */ #if defined(_WIN32) #include #elif defined(__ANDROID__) #include #endif Listing 9-1: Conditional inclusion

Unlike with the if and else keywords, preprocessor conditional inclusion cannot use braces to denote the block of statements controlled by the predicate. Instead, preprocessor conditional inclusion will include all the tokens from the #if, #elif, or #else directive (following the predicate) to the next balanced #elif, #else, or #endif token found, while skipping any tokens in a conditional inclusion branch not taken. Conditional inclusion directives can be nested. You can write #ifdef identifier as shorthand for #if defined identifier or, equivalently, #if defined(identifier). The parentheses around the identifier are optional. Similarly, you can write #ifndef identifier as shorthand for #if !defined identifier. There is no shorthand for #elif defined identifier or #elif !defined identifier.

Generating Errors A conditional inclusion directive may need to generate an error if the preprocessor can’t take any of the conditional branches because no reasonable fallback behavior exists. Consider the example in Listing 9-2, which uses conditional inclusion to select between including the C Standard Library header or the POSIX threading library header . If neither option is available, you should alert the programmer porting the system that the code must be repaired. #if __STDC__ && __STDC_NO_THREADS__ != 1 #include //---snip--#elif POSIX_THREADS == 200809L #include //---snip--#else int compile_error[-1]; // Induce a compilation error #endif Listing 9-2: Inducing a compilation error

Preprocessor

   173

Here, the code generates a diagnostic but doesn’t describe the actual problem. For this reason, C has the #error preprocessing directive, which causes the implementation to produce a diagnostic message. You can optionally follow this directive with one or more preprocessor tokens to include in the resulting diagnostic message. Using these, we can replace the erroneous array declaration from Listing 9-2 with an #error directive such as the one shown in Listing 9-3. #if __STDC__ && __STDC_NO_THREADS__ != 1 #include //---snip--#elif POSIX_THREADS == 200809L #include //---snip--#else #error Neither nor is available #endif Listing 9-3: #error directive

This code generates the following error message if neither threading library header is available: Neither nor is available

Using Header Guards One problem you will face when writing header files is preventing programmers from including the same file twice in a translation unit. Given that you can transitively include header files, you could easily include the same header file multiple times by accident (possibly even leading to infinite recursion between header files). Header guards ensure that a header file is included only once per translation unit. A header guard is a design pattern that conditionally includes the contents of a header file based on whether a header-specific macro is defined. If the macro is not already defined, you define it so that a subsequent test of the header guard will not conditionally include the code. In the program shown in Table 9-4, bar.h uses a header guard (shown in bold font) to prevent its (accidental) duplicate inclusion from foo.c. The first time that "bar.h" is included, the test to see that BAR_H is not defined will return true. We then define the macro BAR_H with an empty replacement list, which is sufficient to define BAR_H, and the function definition for func is included. The second time that "bar.h" is included, the preprocessor will not generate any tokens, because the conditional inclusion test will return false. Consequently, func is defined only once in the resulting translation unit.

174   Chapter 9

Table 9-4: Header Guard

Original sources

Resulting translation unit

bar.h

int func(void) { return 1; }

#ifndef BAR_H #define BAR_H int func(void) { return 1; } #endif /* BAR_H */

int main(void) {  return func();  }

foo.c #include "bar.h"  #include "bar.h" // Repeated inclusion is   // usually not this obvious.    int main(void) {  return func();  }

A common practice when picking the identifier to use as a header file guard is to use the salient parts of the file path, filename, and extension, separated by an underscore and written in all capital letters. For example, if you had a header file that would be included with #include "foo/bar/baz.h", you might choose FOO_BAR_BAZ_H as the header guard identifier. Some IDEs will automatically generate the header guard for you. Avoid using a reserved identifier as your header guard’s identifier, which could introduce undefined behavior. Identifiers that begin with an underscore followed by a capital letter are reserved. For example, _FOO_H is a reserved identifier and a bad choice for a user-chosen header guard identifier, even if you’re including a file named _foo.h. Using a reserved identifier can result in a collision with a macro defined by the implementation, leading to a compilation error or incorrect code.

Macro Definitions The #define preprocessing directive defines a macro. Macros can be used to define constant values or function-like constructs with generic parameters. The macro definition contains a (possibly empty) replacement list—a code pattern that’s injected into the translation unit when the preprocessor expands the macro: #define identifier replacement-list

The #define preprocessing directive is terminated with a newline. In the following example, the replacement list for ARRAY_SIZE is 100.

Preprocessor

   175

#define ARRAY_SIZE 100 int array[ARRAY_SIZE];

In this example, the ARRAY_SIZE identifier is replaced by 100. If no replacement list is specified, the preprocessor will simply remove the macro name. You can typically specify a macro definition on your compiler’s command line, for example, using the -D flag in Clang and GCC or the /D flag in Visual C++. For Clang and GCC, the command line option -DARRAY_SIZE=100 specifies that the macro identifier ARRAY_SIZE is replaced by 100, producing the same result as the #define preprocessing directive from the previous example. If you do not specify the macro replacement list on the command line, compilers will typically provide a replacement list. For example, -DFOO is typically identical to #define FOO 1. The scope of a macro lasts until the preprocessor encounters either an #undef preprocessing directive specifying that macro or the end of the translation unit. Unlike variable or function declarations, a macro’s scope is independent of any block structure. You can use the #define directive to define either an object-like macro or a function-like macro. A function-like macro is parameterized and requires passing a (possibly empty) set of arguments when you invoke it, similar to the way you would invoke a function. Unlike functions, macros let you perform operations using the symbols in the source file. That means you can create a new variable name, or reference the source file and line number the macro is on. An object-like macro is a simple identifier that will be replaced by a code fragment. Table 9-5 illustrates the difference between function-like and objectlike macros. FOO is an object-like macro that is replaced by the tokens (1 + 1) during macro expansion, and BAR is a function-like macro that is replaced by the tokens (1 + (x)), where x is whatever parameter is specified when invoking BAR. Table 9-5: Macro Definition

Original source

Resulting translation unit

#define #define   int i = int j = int k =

int i = (1 + 1);  int j = (1 + (10));  int k = (1 + (2 + 2));

FOO (1 + 1) BAR(x) (1 + (x))  FOO;  BAR(10);  BAR(2 + 2);

The opening parenthesis of a function-like macro definition must immediately follow the macro name, with no intervening whitespace. If a space appears between the macro name and the opening parenthesis, the

176   Chapter 9

parenthesis simply becomes part of the replacement list, as is the case with the object-like FOO macro. The macro replacement list terminates with the first newline character in the macro definition. However, you can join multiple source lines by using the backslash (\) character followed by a newline to make your macro definitions easier to understand. For example, the following definition of the cbrt type-generic macro that computes the cube root of its floating argument #define cbrt(X) _Generic((X), \ long double: cbrtl(X), \ default: cbrt(X), \ float: cbrtf(X) \ )

is equivalent to, but easier to read, than the following: #define cbrt(X) _Generic((X), long double: cbrtl(X), default: cbrt(X), float: cbrtf(X))

One danger when defining a macro is you can no longer use the macro’s identifier in the rest of the program without inducing a macro replacement. For example, as a result of macro expansion, the following invalid program will not compile: #define foo (1 + 1) void foo(int i);

This is because the declaration of foo that the translator receives from the preprocessor is the following invalid declaration: void (1 + 1)(int i);

You can solve this problem by consistently adhering to an idiom throughout your program, such as defining macro names with all uppercase letters, or prefixing all macro names with a mnemonic, as done in some styles of Hungarian notation.12 After you’ve defined a macro, the only way to redefine it is to first invoke the #undef directive for the macro. Once you’ve undefined it, the named identifier no longer represents a macro. For example, the program shown in Table 9-6 defines a function-like macro, includes a header file that uses the macro, and then undefines the macro so that it can be redefined later.

12. H  ungarian notation is an identifier-naming convention in which the name of a variable or function indicates its intention or kind, and in some dialects, its type. Preprocessor

   177

Table 9-6: Undefining Macros

Original sources

Resulting translation unit

header.h

enum Names {  first,  second,  third,  };    void func(enum Names Name) {  switch (Name) {  case first:  case second:  case third:  }  }

NAME(first)  NAME(second)  NAME(third)

file.c enum Names {  #define NAME(X) X,  #include "header.h"  #undef NAME  };    void func(enum Names Name) {  switch (Name) {  #define NAME(X) case X:  #include "header.h"  #undef NAME  }  }

The first use of the NAME macro declares the names of enumerators within the Names enumeration. The NAME macro is undefined and then redefined to generate the case labels in a switch statement. A common idiom is to undefine a macro before redefining it, as shown in Listing 9-4. #undef NAME #define NAME(X) X Listing 9-4: Idiom for safely defining macros

Undefining a macro is safe even when the named identifier isn’t the name of a macro. This macro definition works regardless of whether NAME is already defined.

Macro Replacement Function-like macros may look like functions, but have different behavior. For example, macros let you perform operations using the symbols in the source file. With macros, unlike with functions, you can create new variable names or reference the source file and line number of the macro. When the preprocessor encounters a macro identifier, it will invoke the macro, which expands the identifier to replace it with the tokens from the replacement list, if any, specified in the macro’s definition. For function-like macros, the preprocessor replaces all parameters in the replacement list with the corresponding arguments in the macro invocation after expanding them. Any parameter in the replacement list preceded by a # token is replaced with a string literal preprocessing token

178   Chapter 9

that contains the text of the argument preprocessing tokens (a process sometimes called stringizing). The SRINGIZE macro in Table 9-7 stringizes the value of x. Table 9-7: Stringizing

Original source

Resulting translation unit

#define STRINGIZE(x) #x  const char *str = STRINGIZE(12);

const char *str = "12";

The preprocessor also deletes all instances of the ## preprocessing token in the replacement list, concatenating the preceding preprocessing token with the following token, which is called token pasting. The PASTE macro in Table 9-8 is used to create a new identifier by concatenating foo, the underscore character (_), and bar. Table 9-8: Token Pasting

Original source

Resulting translation unit

#define PASTE(x, y) x ## _ ## y  int PASTE(foo, bar) = 12;

int foo_bar = 12;

After expanding the macro, the preprocessor rescans the replacement list to expand additional macros within it. If the preprocessor finds the name of the macro being expanded while rescanning—including the rescanning of nested macro expansions within the replacement list—it won’t expand the name again. Furthermore, if macro expansion results in a fragment of program text that’s identical to a preprocessing directive, that ­f ragment won’t be treated as a preprocessing directive. During macro expansion, a repeated parameter name in the replacement list will be replaced multiple times by the argument given in the invocation. This can have surprising effects if the argument to the macro invocation involves side effects, as shown in Table 9-9. This problem is explained in detail in CERT C rule PRE31-C (Avoid side effects in arguments to unsafe macros). Table 9-9: Unsafe Macro Expansion

Original source

Resulting translation unit

#define bad_abs(x) (x >= 0 ? x : -x)    int func(int i) {  return bad_abs(i++);  }

int func(int i) {  return (i++ >= 0 ? i++ : -i++);  }

In the macro definition in Table 9-9, each instance of the macro parameter x is replaced by the macro invocation argument i++, causing i to be incremented twice in a way that a programmer or reviewer reading the

Preprocessor

   179

original source code can easily overlook. Parameters like x in the replacement list, as well as the replacement list itself, should usually be fully parenthesized as in ((x) >= 0 ? (x) : -(x)) to prevent portions of the argument x from associating with other elements of the replacement list in an unexpected way. Another potential surprise is that a comma in a function-like macro invocation is always interpreted as a macro argument delimiter. The C Standard ATOMIC_VAR_INIT macro, which allows you to initialize an arbitrary atomic variable, demonstrates the danger (Table 9-10). This code fails to translate because the comma in ATOMIC_VAR_INIT({1, 2}) is treated as a f­ unction-like macro argument delimiter, causing the preprocessor to interpret the macro as having two syntactically invalid arguments {1 and 2} instead of a single, valid argument {1, 2}.13 Table 9-10: ATOMIC_VAR_INIT Macro

Original sources

Resulting translation unit

stdatomic.h

#define ATOMIC_VAR_INIT(value) (value)

foo.c #include     struct S {  int x, y;  };    _Atomic struct S val = ATOMIC_VAR_INIT({1, 2});

Type-Generic Macros The C programming language doesn’t allow you to overload functions based on the types of the parameters passed to the function, as you can in other languages such as Java and C++. However, you might sometimes need to alter the behavior of an algorithm based on the types of the arguments involved. For example, has three sin functions (sin, sinf, and sinl) because each of the three floating-point types (double, float, and long ­double, respectively) has a different precision. Using generic selection expressions, you can define a single function-like identifier that delegates to the correct underlying implementation based on the argument type when called. A generic selection expression maps the type of its unevaluated operand expression to an associated expression. If none of the associated types match, it can optionally map to a default expression. You can use type-generic macros (macros that include generic selection expressions) to make your code more readable. In Table 9-11, we define a type-generic macro to select the correct variant of the sin function from .

13. This usability issue is one of the reasons the ATOMIC_VAR_INIT macro was deprecated in C17.

180   Chapter 9

Table 9-11: Generic Selection Expression as a Macro

Original source #define sin(X) float: sinf, double: sin, long double: )(X)

Resulting _Generic resolution _Generic((X), \  \  \  sinl \ 

int main(void) {  float f = sinf(1.5708f);  double d = sin(3.14159);  }

int main(void) {  float f = sin(1.5708f); double d = sin(3.14159);  }

The controlling expression (X) of the generic selection expression is unevaluated; the type of the expression selects a function from the list of type : expr mappings. The generic selection expression picks one of these function designators (either sinf, sin, or sinl) and then executes it. In this example, the argument type in the first call to sin is float so the generic selection resolves to sinf, and the argument type in the second call to sin is double so this resolves to sin. Because this generic selection expression has no default association, an error occurs if the type of (X) doesn’t match any of the associated types. If you include a default association for a generic selection expression, it will match every type not already used as an association, including types you might not expect, such as pointers or structure types. Type-generic macro expansion can be difficult to use when the resulting value type depends on the type of an argument to the macro, as with the sin example in Table 9-11. For example, it would be a mistake to call the sin macro and assign the result to an object of a specific type, or pass its result as an argument to printf, because the necessary object type or format specifier will depend on whether sin, sinf, or sinl is called. Examples of type-generic macros for math functions can be found in the C Standard Library header.

Predefined Macros Some macros are defined automatically by the implementation without requiring you to include a header file. These macros are called predefined macros because they’re implicitly defined by the preprocessor rather than explicitly defined by the programmer. For example, the C Standard defines numerous macros that you can use to interrogate the compilation environment or provide basic functionality. Some other aspects of the implementation (such as the compiler or the compilation target operating system) also automatically define macros. Table 9-12 lists some of the common macros defined by the C Standard. You can obtain a full list of predefined macros from Clang or GCC by passing the -E -dM flags to these compilers. You should check your implementation’s documentation for more information.

Preprocessor

   181

Table 9-12: Predefined Macros

182   Chapter 9

Macro name

Replacement and purpose

__DATE__

A string literal of the date of translation of the preprocessing translation unit in the form Mmm dd yyyy.

__TIME__

A string literal of the time of translation for the preprocessing translation unit in the form hh:mm:ss.

__FILE__

A string literal representing the presumed filename of the current source file.

__LINE__

An integer constant representing the presumed line number of the current source line.

__STDC__

The integer constant 1 if the implementation conforms to the C Standard.

__STDC_HOSTED__

The integer constant 1 if the implementation is a hosted implementation, or the integer constant 0 if it is stand-alone. This macro is conditionally defined by the implementation.

__STDC_VERSION__

The integer constant representing the version of the C Standard the compiler is targeting, such as 201710L for the C17 standard.

__STDC_ISO_10646__

An integer constant of the form yyyymmL. This macro is conditionally defined by the implementation. If this symbol is defined, every character in the Unicode required set, when stored in an object of type wchar_t, has the same value as the short identifier of that character.

__STDC_UTF_16__

The integer constant 1 if values of type char16_t are UTF-16 encoded. This macro is conditionally defined by the implementation.

__STDC_UTF_32__

The integer constant 1 if values of type char32_t are UTF-32 encoded. This macro is conditionally defined by the implementation.

__STDC_NO_ATOMICS__

The integer constant 1 if the implementation doesn’t support atomic types, including the _Atomic type qualifier, and the header. This macro is conditionally defined by the implementation.

__STDC_NO_COMPLEX__

The integer constant 1 if the implementation doesn’t support complex types or the header. This macro is conditionally defined by the implementation.

__STDC_NO_THREADS__

The integer constant 1 if the implementation doesn’t support the header. This macro is conditionally defined by the implementation.

__STDC_NO_VLA__

The integer constant 1 if the implementation doesn’t support variable-length arrays or variably modified types. This macro is conditionally defined by the implementation.

Summary In this chapter, you learned about some of the features provided by the preprocessor. You learned how to include fragments of program text in a translation unit, conditionally compile code, and generate diagnostics on demand. You then learned how to define and undefine macros, how macros are invoked, and about macros that are predefined by the implementation. In the next chapter, you’ll learn how to structure your program into more than one translation unit to create more maintainable programs.

Preprocessor

   183

10 PROGR AM STRUCTURE with Aaron Ballman

Any real-world system is made up of multiple components, such as source files, header files, and libraries. Many also contain resources including images, sounds, and configuration files. Composing a program from smaller logical components is a good software engineering practice, because these components are easier to manage than a single large file. In this chapter, you’ll learn how to structure your program into multiple units consisting of both source and include files. You’ll also learn how to link multiple object files together to create libraries and executable files.

Principles of Componentization Nothing prevents you from writing your entire program within the main function of a single source file. However, as the function grows in size, this approach will quickly become unmanageable. For this reason, it makes sense to decompose your program into a collection of components that exchange information across a shared boundary, or interface. Organizing source code into components makes it easier to understand, and it allows you to reuse the code elsewhere in the program, or even with other programs. Understanding how best to decompose a program typically requires experience. Many of the decisions programmers make are driven by performance. For example, you may need to minimize communication over a high-latency interface. Or, a client may handle input field validation from the user interface so that it doesn’t require a round trip to the server. In this section, we’ll cover some principles of component-based software engineering.

Coupling and Cohesion In addition to performance, the aim of a well-structured program is to achieve desirable properties like low coupling and high cohesion. Cohesion is a measure of the commonality between elements of a programming interface. Assume, for example, that a header file exposes functions for calculating the length of a string, calculating the tangent of a given input value, and creating a thread. This header file has low cohesion, because the exposed functions are unrelated to each other. Conversely, a header file that exposes functions to calculate the length of a string, concatenate two strings together, and search for a substring within a string has high cohesion, because all of the functionality is related. This way, if you need to work with strings, you need only to include the string header file. Similarly, related functions and type definitions that form a public interface should be exposed by the same header file to provide a highly cohesive interface of limited functionality. We’ll discuss public interfaces further in “Data Abstractions” on page 187. Coupling is a measure of the interdependency of programming interfaces. For example, a tightly coupled header file can’t be included in a program by itself; instead, it must be included with other header files in a specific order. You may couple interfaces for a variety of reasons, such as a mutual reliance on data structures, interdependence between functions, or the use of a shared global state. But when interfaces are tightly coupled, modifying program behavior becomes difficult, because changes can have a ripple effect across the system. You should always strive to have loose coupling between interface components, regardless of whether they’re members of a public interface or implementation details of the program. By separating your program logic into distinct, highly cohesive components, you make it easier to reason about the components and test the program (because you can verify the correctness of each component independently). The end result is a more maintainable, less buggy system. 186   Chapter 10

Code Reuse Code reuse is the practice of implementing functionality only once, and then reusing it in various parts of the program without duplicating the code. Code duplication can lead to subtly unexpected behavior, oversized and bloated executables, and increased maintenance costs. And anyway, why write the same code more than once? Functions are the lowest-level reusable units of functionality. Any logic that you might repeat more than once is a candidate for encapsulating in a function. If the functionality has only minor differences, you can frequently parameterize the function to serve multiple purposes. Each function should perform work that isn’t duplicated by any other function. You can then compose individual functions to solve increasingly sophisticated problems. Packaging reusable logic into functions can improve maintainability and eliminate defects. For example, though you could determine the length of a null-terminated string by writing a simple for loop, it’s more maintainable to use the strlen function from the C Standard Library. Because other programmers are already familiar with the strlen function, they’ll have an easier time understanding what that function is doing than what the for loop is doing. Furthermore, if you reuse existing functionality, you’re less likely to introduce behavioral differences compared to ad hoc implementations, and you make it easier to globally replace the functionality with a better-performing algorithm or more secure implementation. When designing functional interfaces, a balance must be struck between generality and specificity. An interface that’s specific to the current requirement may be lean and effective, but hard to modify when requirements change. A general interface might allow for future requirements, but be cumbersome for foreseeable needs.

Data Abstractions A data abstraction is any reusable software component that enforces a clear separation between the abstraction’s public interface and the implementation details. The public interface for each data abstraction includes the data-type definitions, function declarations, and constant definitions required by the users of the data abstraction and is placed in header files. The implementation details of how the data abstraction is coded, as well as any private utility functions, are hidden within a source file, or a header file that’s in a separate location from the public interface header files. This separation of the public interface from the private implementation allows you to change the implementation details without breaking code that depends on your component. Header files typically contain function declarations and type definitions for the component. For example, the C Standard Library provides the public interface for string-related functionality, while provides utility functions for threading. This logical separation has low coupling and high cohesion. This approach makes it easier to access only the specific components you need, and reduces compile times and the likelihood of name collisions. You don’t need to know anything about threading APIs, for example, if all you need is the strlen function. Program Structure

   187

Another consideration is if you should explicitly include the headers required by your header file, or require the users of the header file to include them first. In terms of producing data abstractions, it is a good idea for headers to be self-contained and include the headers they use. Not doing so is more of a burden on the users of the abstraction, and also leaks implementation details about the data abstraction. Examples in this book do not always follow this practice, in an attempt to keep these file listings concise. Source files implement the functionality declared by a given header file or the application-specific program logic used to perform whatever actions are needed for a given program. For example, if you have a network.h header file that describes a public interface for network communications, you may have a network.c source file (or network_win32.c for Windows only and network_linux.c for Linux only) that implements the network communication logic. It’s possible to share implementation details between two source files by using a header file, but the header file should be placed in a distinct location from the public interface so as not to accidentally expose implemen­ tation details. A collection is a good example of a data abstraction that separates the basic functionality from the implementation or underlying data structure. A collection represents a group of data elements and supports operations such as adding elements to the collection, removing data elements from the collection, and checking to see if the collection contains a specific data element. There are many ways to implement a collection. For example, a collection of data elements may be represented as a flat array, a binary tree, a directed (possibly acyclic) graph, or a different structure. The choice of data structure can impact an algorithm’s performance, depending on what kind of data you’re representing and how much data there is to represent. For example, a binary tree may be a better abstraction for a large amount of data that needs good lookup performance, whereas a flat array is likely a better abstraction for a small amount of data of fixed size. Separating the interface of the collection data abstraction from the implementation of the underlying data structure allows the implementation to change without requiring changes to code that relies on the collection interface.

Opaque Types Data abstractions are most effective when used with opaque data types that hide information. In C, opaque (or private) data types are those expressed using an incomplete type, such as a forward-declared structure type. An incomplete type is a type that describes an identifier but lacks information needed to determine the size of objects of that type or their layout. Hiding internal-only data structures discourages programmers who use the data abstraction from writing code that depends on implementation details, which may change. The incomplete type is exposed to users of the data abstraction, while the fully defined type is accessible only to the implementation. Let’s say that we want to implement a collection that supports a limited number of operations, such as adding an element, removing an element, 188   Chapter 10

and searching for an element. The following example implements collection  _type as an opaque type, hiding the implementation details of the data type from the library’s user. To accomplish this, we create two header files: an external collection.h header file included by the user of the data type, and an internal file included only in files that implement the functionality of the data type. In the external collection.h header file, the collection_type data type is defined as an instance of struct collection_type, which is an incomplete type: typedef struct collection_type collection_type; // Function declarations extern errno_t create_collection(collection_type **result); extern void destroy_collection(collection_type *col); extern errno_t add_to_collection(collection_type *col, const void *data, size_t byteCount); extern errno_t remove_from_collection(collection_type *col, const void *data, size_t byteCount); extern errno_t find_in_collection(const collection_type *col, const void *data,   size_t byteCount); //---snip---

The collection_type identifier is aliased to struct collection_type (an incomplete type). Consequently, functions in the public interface must accept a pointer to this type, instead of an actual value type, because of the constraints placed on the use of incomplete types in C. In the internal header file, struct collection_type is fully defined but not visible to a user of the data abstraction: struct node_type { void *data; size_t size; struct node_type *next; }; struct collection_type { size_t num_elements; struct node_type *head; };

Modules that implement the abstract data type include both the external and internal definitions, whereas users of the data abstraction include only the external collection.h file. This allows the implementation of the c­ ollection_type data type to remain private.

Executables In Chapter 9, you learned that the compiler is a pipeline of translation phases, and that the compiler’s ultimate output is object code. The last phase of translation, called the link phase, takes the object code for all of the translation units in the program and links it together to form a final executable. This can be an executable that a user can run, such as a.out or foo.exe, a library, or a more specialized program such as a device driver Program Structure

   189

or a firmware image (machine code to be burned onto a ROM). Linking allows you to separate your code into distinct source files that can be compiled independently, which helps you build reusable components. Libraries are executable components that cannot be executed independently. Instead, you can incorporate a library into executable programs. You can invoke the functionality of the library by including the library’s header files in your source code and calling the declared functions. The C Standard Library is an example of a library—you include the header files from the library, but you do not directly compile the source code that implements the library functionality. Instead, the implementation ships with a prebuilt version of the library code. Libraries allow you to build on the work of others for the generic components of a program so you can focus on developing logic that is unique to your program. For example, when writing a video game, reusing existing libraries should allow you to focus on developing the game logic, not worrying about the details of retrieving user input, network communications, or graphics rendering. Libraries can often allow a program written with one compiler to use code that was built by a different compiler. Libraries are linked into your application and can be either static or dynamic. A static library, also known as an archive, incorporates its machine or object code directly into the resulting executable, which means that a static library is often tied to a specific release of the program. Because a static library is incorporated at link time, the contents of the static library can be further optimized for your program’s use of the library. Library code used by the program is available for link-time optimizations, while unused library code can be stripped from the final executable. A dynamic library, also referred to as a shared library or a dynamic shared object, is an executable without the startup routines. It can be packaged with the executable or installed separately but must be available when the executable calls a function provided by the dynamic library. Many modern operating systems will load the dynamic library code into memory once and share it across all the applications that need it. You can replace a dynamic library with different versions as necessary after your application has been deployed. Letting the library evolve separately from the program comes with its own set of benefits and risks. A developer can correct bugs in the library after an application has already shipped without requiring the application to be recompiled, for instance. However, dynamic libraries provide a potential opportunity for a malicious attacker to replace a library with a nefarious one or an end user to accidentally use an incorrect version of the library. It is also possible to make a breaking change in a new library release that results in an incompatibility with existing applications that use the library. Static libraries might execute somewhat faster because its object code (binary) is already included in the executable file. Generally speaking, the benefits of using dynamic libraries outweigh the disadvantages. Each library has one or more header files that contain the public interface to the library, and one or more source files that implement the logic for the library. You can benefit from structuring your code as a collection of libraries even if the components aren’t turned into actual libraries. Using 190   Chapter 10

an actual library makes it harder to accidentally design a tightly coupled interface where one component has special knowledge of the internal details of another component.

Linkage Linkage is a process that controls whether an interface is public or private and determines whether any two identifiers refer to the same entity. C provides three kinds of linkage: external, internal, or none. When a declaration has external linkage, identifiers referring to that declaration all refer to the same entity (such as a function or an object) everywhere in the program. When a declaration has internal linkage, identifiers referring to that declaration refer to the same entity only within the translation unit containing the declaration. If two translation units both refer to the same internal linkage identifier, they refer to different instances of the entity. If a declaration has no linkage, it is a unique entity in each translation unit. The linkage of a declaration is either explicitly declared or implicitly implied. If you declare an entity at file scope without explicitly specifying extern or static, the entity is implicitly given external linkage. Identifiers that have no linkage include function parameters, block scope identifiers declared without an extern storage class specifier, or enumeration constants. Listing 10-1 shows examples of declarations of each kind of linkage. static int i; // i is declared with explicit internal linkage. extern void foo(int j) { // foo is declared with explicit external linkage. // j has no linkage because it is a parameter. } Listing 10-1: Examples of external, internal, and no linkage

If you explicitly declare an identifier with the static storage class specifier at file scope, it has internal linkage. The static keyword gives internal linkage only to file scope entities. Declaring a variable at block scope as static creates an identifier with no linkage, but it does give the variable static storage duration. As a reminder, static storage duration means its lifetime is the entire execution of the program and its stored value is initialized only once, prior to program startup. The different meanings of static when used in different contexts is obviously confusing and consequently a common interview question. You can create an identifier with external linkage by declaring it with the extern storage class specifier. This works only if you haven’t previously declared the linkage for that identifier. The extern storage class specifier has no effect if a prior declaration gave the identifier linkage. Declarations with conflicting linkage can lead to undefined behavior; see CERT C rule DCL36-C (Do not declare an identifier with conflicting linkage classifications) for more information. Table 10-1 shows examples of declarations with explicit and implicit linkage. Program Structure

   191

Table 10-1: Examples of Implicit and Explicit Linkage

Implicit and explicit linkage foo.c void func(int i) { // Implicit external linkage  // i has no linkage }  static void bar(void); // Internal linkage, different bar than bar.c extern void bar(void) { // bar still has internal linkage because the initial declaration was // declared as static; the extern specifier has no effect in this case. }

bar.c extern void func(int i); // Explicit external linkage static void bar(void) { // Internal linkage; different bar than foo.c func(12); // Calls func from foo.c }  int i; // External linkage; does not conflict with i from foo.c or bar.c void baz(int k) { // Implicit external linkage bar(); // Calls bar from bar.c, not foo.c }

The identifiers in your public interface should have external linkage so that they can be called from outside their translation unit, while the identifiers that are implementation details should be declared with internal or no linkage. A common approach to achieving this is to declare your public interface functions in a header file with or without using the extern storage class specifier (the declarations implicitly have external linkage, but there is no harm in explicitly declaring them with extern), and define the public interface functions in a source file in a similar manner. However, within the source file, all declarations that are implementation details should be explicitly declared static to keep them private, accessible to just that source file. You can include the public interface declared within the header file by using the #include preprocessor directive to access its interface from another file. A good rule of thumb is that file-scope entities that don’t need to be visible outside the file should be declared as static. This practice limits the pollution of the global namespace and decreases the chances of surprising interactions between translation units.

Structuring a Simple Program To learn how to structure a complex, real-world program, let’s develop a simple program to determine whether a number is prime. A prime number (or a prime) is a natural number greater than one that cannot be formed by multiplying two smaller natural numbers. We’ll write two separate components: a static library containing the testing functionality and a command line application that provides a user interface for the library.

192   Chapter 10

The primetest program accepts a whitespace-delimited list of integer values as input and outputs whether each value is a prime number. If any of the inputs are invalid, the program will output a helpful message explaining how to use the interface. Before exploring how to structure the program, let’s examine the user interface. First, we print the help text for the command line program, as shown in Listing 10-2. // Print command line help text. static void print_help(void) { printf("%s", "primetest num1 [num2 num3 ... numN]\n\n"); printf("%s", "Tests positive integers for primality. Supports testing "); printf("%s [2-%llu].\n", "numbers in the range", ULLONG_MAX); } Listing 10-2: Printing help text

The print_help function consists of three separate calls to the printf function that print help text about how to use the command to the standard output. Next, because the command line arguments are passed to the program as textual input, we define a utility function to convert them to integer ­values, as shown in Listing 10-3. // Converts a string argument arg to an unsigned long long value referenced by val. // Returns true if the argument conversion succeeds, and false if it fails. static bool convert_arg(const char *arg, unsigned long long *val) { char *end; // strtoll returns an in-band error indicator; clear errno before the call. errno = 0; *val = strtoull(arg, &end, 10); // if if if

Check for failures where the call returns a sentinel value and sets errno. ((*val == ULLONG_MAX) && errno) return false; (*val == 0 && errno) return false; (end == arg) return false;

// If we got here, we were able to convert the argument. However, we want to allow only // values greater than one, so we reject values